361 results about "Graph encoding" patented technology

The invention belongs to the technical field of network security, provides a network attack detection method based on dynamic graph coding, and solves the problems of poor dynamic adaptability of an attack path and missing of timing constraint in the prior art. The method comprises the following steps: constructing a dynamic threat map, extracting a triple of heterogeneous threat intelligence by using a RoBERTa model, and adding a timestamp and a confidence attribute; a dynamic graph encoder for time sequence perception is designed, semantic and evolution laws are fused through periodic time coding and a multi-head time sequence attention mechanism, and feature weights are adjusted in combination with a gating residual layer; an event-driven incremental updating strategy is adopted, and node similarity is calculated to achieve local subgraph updating; a time sequence rule base is established, three-dimensional parameter verification attack chain time sequence logic is defined, and abnormity is judged through conflict scores; and finally, integrating a graph updating module, a dynamic coding module and a constraint analysis module to realize multi-source threat feature matching and attack detection. According to the method, the adaptability of attack path evolution is improved through dynamic graph modeling and real-time increment updating.

The invention relates to the technical field of big language models, and discloses a big language model training method based on knowledge graph enhancement, comprising the following steps: S1, constructing a multi-source heterogeneous knowledge graph; s2, coding the mixed attention heterogeneity map; s3, bidirectionally mapping a pre-training task; and S4, position specific gating fusion. According to the big language model training method and system based on knowledge graph enhancement, a same proton graph is established for a structured triple and text entity description, nodes are connected across graph edges to form a heterogeneous graph, associated edges are established through entity linking and syntactic analysis, multi-source knowledge is modeled in a unified mode, and the problem of low fusion efficiency is solved; mixed attention coding adopts a layering mechanism, a semantic level calculates weights according to type compatibility, a node level calculates similarity aggregation features through cosine distance and path length, entity vectors are generated through pooling, map structures and semantics are explicitly learned, reasoning accuracy is improved, and the problem of knowledge understanding superficial layer is solved.

The invention belongs to the field of bioinformatics, and relates to a molecular property prediction method based on multi-modal gating and comparative learning, which comprises the technologies of comparative learning, graph neural network, cross-modal alignment, gating attention and the like. Firstly, data standardization and graph construction are carried out, and molecular fingerprint embedding is extracted; secondly, a heterogeneous dual-channel graph coding architecture is adopted, one path captures atom short-range interaction through an attention mechanism, the other path integrates a molecular global structure and long-range dependence, and complementary molecular representation is generated; then, a cross-modal attention mechanism is introduced, bidirectional association of graph and fingerprint features is achieved, and modal weights are adaptively and dynamically distributed through a gating fusion module; and finally, a comparison pre-training strategy is adopted, a molecular graph and fingerprints are utilized to construct a sample pair, and discriminative molecular representation is learned on unlabeled data. According to the method, the accuracy of molecular property prediction is remarkably improved, and an efficient and reliable calculation tool is provided for virtual drug screening and lead compound optimization.

The invention relates to the technical field of computer-generated holographic three-dimensional display and deep learning, in particular to an unsupervised deep learning method for realizing three-dimensional holographic display. The method comprises the steps of generating a depth map corresponding to a two-dimensional image; the depth image and the two-dimensional image are spliced in the channel dimension to serve as input of a hologram encoder, and a double-U-Net cascade neural network architecture serves as the hologram encoder; angular spectrum diffraction back propagation is carried out through the generated pure phase hologram to realize three-dimensional scene discretization reconstruction, a reconstructed image with a specified depth is obtained, a depth map is uniformly quantized to obtain a plurality of binary masks, loss calculation is carried out on the reconstructed image and a target image superposed with the corresponding depth binary masks, network parameter optimization is carried out, and a target image with the depth corresponding to the target image is obtained. And when the training of the double U-Net cascade neural network architecture is converged, the training stage is ended. According to the method, high-quality three-dimensional hologram reconstruction is realized through layered angular spectrum propagation, and the method has relatively high precision and detail reduction capability.

The invention provides a geological data driving path optimization and settlement prediction method for pipe jacking construction, and relates to the technical field of artificial intelligence. The method comprises the following steps: firstly, acquiring sparse drilling point location geological data, constructing a geological data set including soil layer types, forming continuous geological feature tensors through multi-dimensional interpolation, and generating a path vector sequence by combining starting and ending points and building distribution to represent crossing tracks of different paths under geological conditions; and constructing a path graph topological structure through a graph coding neural network, performing feature propagation, obtaining a path comprehensive score vector, and screening an optimal path meeting structural integrity and settlement response constraints. Further identifying settlement trend sensitive points by utilizing the settlement risk prediction values and the accessibility scores of the nodes, constructing a settlement trend map and calculating settlement transaction coefficients. And the transaction coefficient is fed back to a path generation link and is used for iteratively optimizing path sampling density and node distribution so as to realize dynamic correction and stable convergence of a path scheme.

The invention provides a real-time video image compression method based on deep learning, and relates to the technical field of video image compression, and the method comprises the steps: carrying out the key feature recognition through employing an attention mechanism; performing convolution training optimization on the video image sample data set by using a deep learning network structure; a self-encoder structure is designed to carry out feature map encoding compression; a video image compression adaptive network is generated through series fusion; a real-time video image frame is collected for preprocessing, and feature compression processing is performed on a standard video image frame based on a video image compression adaptive network. According to the method and the device, the technical problem that the video compression quality is reduced due to the fact that the generalization ability is insufficient in the face of various scenes and the video compression strategy is difficult to adaptively adjust according to different scenes in the prior art can be solved, the adaptive network is constructed through the combination of deep learning and the auto-encoder, and the video compression quality is improved. And the video compression strategy is dynamically adjusted according to the contents of different video images, so that the video compression quality is improved.

The invention relates to the technical field of slope catastrophe intelligent early warning and data modeling, and particularly discloses a strip mine high and steep slope multi-source monitoring and early warning method based on a graph neural network and Transform, and the method comprises the following steps: S01, carrying out the data preprocessing and disturbance variable construction of monitoring data; s02, constructing a heterogeneous space diagram structure by taking the monitoring points as nodes and taking geography, lithology and dynamic response relationships as edges; s03, constructing a space-time end-to-end multilayer coding framework based on the graph attention network and the integrated deep neural structure; s04, on the basis of graph coding and time sequence output, introducing a disturbance variable embedding mechanism, and designing a joint attention fusion structure; and S05, generating a deformation trend prediction value of the slope in a future period of time and performing corresponding risk grade judgment. The invention aims to solve the key technical problem of weak adaptability and interpretability of an early warning system.

The invention provides a method and equipment for constructing a text entity recognition model based on large model data enhancement. The method comprises the following steps: firstly, constructing an initial model comprising a preprocessing unit, a syntax dependency analysis unit, a data correction enhancement unit, a context coding unit, a syntax enhancement unit, an expression fusion unit and a sequence decoding processing unit; preprocessing the training sample to obtain a preprocessed text, and establishing a preliminary dependency graph through syntactic analysis; correcting and enhancing the text and the dependency graph by the large language model to obtain a text sequence and a dependency graph for subsequent use; encoding the text sequence to obtain an initial representation vector, and enhancing the vector in combination with the dependency graph; after fusion, decoding and outputting a prediction label containing a lexical entity label; and calculating loss by using a function containing conditional random field structure loss, judging convergence, and if not, updating parameters and continuing training until a target model is obtained. According to the method, data are optimized and expanded by means of a large language model, syntactic dependency enhancement features are combined, and the recognition capability of the complex text named entities is improved.

The invention discloses a copywriting generation method based on artificial intelligence. The method comprises the following steps: constructing a propagation technique knowledge base and a creation technique knowledge base of an advertisement copywriting according to demand information; retrieving a matched creative strategy combination from the creative technique knowledge base, and encoding the creative strategy combination as a first cue word fragment; outputting a Top-K sub-graph, and encoding the Top-K sub-graph into a second cue word fragment; splicing the first cue word segment, the second cue word segment and the target audience portrait and style control parameters, generating a complete cue word, and generating an advertisement copywriting first draft; performing quality evaluation and generating a hallucination probability graph; and outputting the finally generated advertisement copywriting until the illusion probability output by the rewritten advertisement copywriting meets the requirement. According to the method, a closed loop of copywriting generation-evaluation-optimization is realized, and the quality of the copywriting can be effectively improved through a small number of iterations.

The invention provides a ship abnormal behavior detection method based on an improved graph attention neural network, and belongs to the field of ship abnormal behavior detection. Through AIS data, a graph structure including attribute features such as ship position, speed and course is constructed, attribute features of ship tracks are expanded by using dynamic cavity graph convolution, and a space-time attention mechanism is fused into a multi-head attention mechanism, so that the model has the capability of processing time dimension features to better extract space-time features of the ship tracks, and the time-space attention mechanism is optimized. The detection effect of the abnormal behavior of the single ship is improved; in the aspect of abnormal behavior detection between ships, a two-layer graph coding structure is constructed, the first-layer graph coding adopts a coding mode of single ship detection to obtain trajectory features of the ships, and the second-layer graph coding constructs relation features between ship trajectories based on the trajectory features of each ship. Effective detection of abnormal behaviors between ships is realized in combination with the attention mechanism of the GAT, and the method has good accuracy and robustness and high practical value.

The invention discloses a human body grid reconstruction method based on Transform, and the method comprises the steps: firstly obtaining a public human body data set, and carrying out the standardization of an image in the human body data set; secondly, images in the human body data set generate a plurality of visual angle features through two branches of a front view encoder and a visual angle conversion network respectively, optimization and enhancement are carried out respectively, the optimized and enhanced features are fused, and a multi-visual angle aggregation feature is generated; and then, in combination with the optimization of a Transform encoder, initial joint features are extracted from the images in the human body data set, and final posture features are generated. And finally, constructing a grid regression module, adding the multi-view aggregated features and the attitude features, and inputting the added features to the grid regression module to generate a final human body grid reconstruction result. According to the method, the feature information of different visual angles is fused, so that the precision and detail representation of human body grid reconstruction are effectively improved under the conditions of complex posture change and shielding.

The invention belongs to the field of medical vision-language pre-training, and relates to a medical vision-language pre-training method based on multi-view and text combination, which comprises the following steps of: preprocessing chest X-ray image-report data; inputting the preprocessed image into a view encoder to obtain local features VF and VL and global representations gF and gL of a positive view and a side view, wherein the local features VF and VL of the positive view and the local features VL of the side view and the global representations gF and gL of the positive view and the side view are compared with the report sequence XT of the mask and the report sequence # imgabs0 # of the mask; respectively inputting the sequence XT and # imgabs1 # into a report encoder to obtain a local feature T, a global representation gT and a mask report representation # imgabs2 #, and inputting VF, VL, T, gF, gL, gT and # imgabs3 # into a positive side view feature integration module to obtain a mask report generation text T'and a prediction probability P thereof; inputting the VF, the VL and the T into a positive and lateral feature alignment module to obtain fine-grained representations F and L of positive and lateral views; calculating a loss function value according to P, F, L, gF, gL and gT, and updating model parameters according to the loss function value until a pre-trained medical vision-language general model is obtained; according to the method, the lateral view is introduced into pre-training, so that the diagnosis accuracy is improved.

The invention relates to a power distribution network multi-source heterogeneous data fusion processing method, and belongs to the technical field of power system data processing, and the method comprises the following steps: carrying out the preprocessing of multi-source heterogeneous data collected by a power distribution network monitoring device, and generating a two-dimensional data matrix; extracting spatial topological features through a graph neural network in a space-time graph encoder, capturing time dynamic features in combination with a long-short-term memory network, and outputting a hidden state vector fused with space-time features; after the vector is input into a multi-criterion generator, an evaluation criterion is generated by a plurality of criterion sub-modules; in the process, the adversarial training module dynamically optimizes a criterion to generate a threshold parameter through a generator-discriminator architecture to form a closed-loop feedback mechanism; and finally, the dynamic output layer comprehensively optimizes the criterion, and outputs a state evaluation report including operation state evaluation, accurate fault diagnosis and a resource optimization configuration scheme through space-time correlation feature cross analysis.

The invention discloses a 3D digital human facial expression synthesis method based on audio driving and a visualization system. The method comprises the following steps: acquiring a data set; inputting the original audio in the data set into an audio encoder in the model, and extracting audio features; inputting the audio features into a KAN-based decoder in the model to generate 3D digital human facial expression actions; inputting 3D digital human facial expression actions into a graph encoder in the model, and extracting lip features; collaborative optimization and model training of multi-modal features are realized by constructing a joint loss function fusing audio features, 3D facial expression parameters and lip features; and inputting an original audio to be tested into the trained model, and generating corresponding 3D digital human facial expression actions through the audio encoder and the KAN decoder in sequence. According to the invention, the generated 3D digital human facial expression action is more vivid and closer to the real character expression.

The embodiment of the invention provides an image recognition method based on a visual language, a controller, a robot and a medium, relates to the technical field of artificial intelligence, and is suitable for the fields of financial science and technology and medical health. The method comprises the following steps: acquiring article use record information of a target user; generating an auxiliary object searching knowledge graph based on the object use record information; carrying out graph coding on the auxiliary object searching knowledge graph to obtain graph structured semantic features; collecting a scene where the target user is located to obtain a current scene picture, and performing visual coding on the current scene picture to obtain current picture features; carrying out attention fusion processing on the map structured semantic features and the current picture features to obtain map picture joint features; and obtaining an object searching natural language instruction containing the target object description information, and determining an object searching condition according to the object searching natural language instruction and the map picture joint features. According to the embodiment of the invention, the article searching accuracy in a specific scene can be improved.

The invention relates to an APT attack detection and machine learning technology, in particular to an APT attack detection method fusing comparative learning and cross-domain recommendation. A traceability graph is constructed based on a data set, and a bipartite graph is constructed according to the traceability graph; generating an r-ego network for all nodes in the bipartite graph; generating positive and negative sample pre-training graph encoders in the source domain, and transferring the pre-trained graph encoders to the target domain; generating initialization embedding in a target domain by using a pre-trained graph encoder; the matrix decomposition model is finely adjusted by using the initialized embedding, and the final embedding of each node in the target domain is obtained by using the finely-adjusted matrix decomposition model; and predicting whether the two nodes interact based on the final embedding. According to the method, the high-order connectivity is formed by using the side information of the system entities to predict the possibility of interaction between the entities. And meanwhile, a cross-domain recommendation method is used, so that the problem of data sparsity of APT attacks is relieved while the recommendation performance on a target domain is improved.

The invention relates to the technical field of automatic driving risk analysis, in particular to a traffic risk analysis method and analysis device based on a graph neural network, and the method comprises the steps: sequentially carrying out the coding of a graph encoder and the coding of an attention mechanism based on the obtained historical track information of an intelligent agent and the lane line information of a map, and obtaining an attention mechanism; obtaining an attention mechanism coding result; decoding by using a cross attention mechanism based on an attention mechanism coding result; and performing prediction trajectory decoding by using a multi-layer perceptron based on a cross attention mechanism decoding result, performing decoding by using a deconvolution network based on a decoding result, calculating risk matrix loss by using a probability loss function, and performing model training and verification to obtain a risk analysis model based on a graph neural network so as to predict traffic risks. Therefore, the problems that the risk analysis depth of traffic participants is limited and future intentions and potential risks of the traffic participants are difficult to accurately describe by traffic risk analysis methods in related technologies are solved.

The invention discloses a multi-language code generation method based on self-supervised pre-training, which comprises the following steps: acquiring and cleaning multi-language code data to form a training corpus; the method comprises the following steps: representing code data as an abstract syntax tree, extracting a control flow diagram and a data flow diagram of the code data, and obtaining unified semantic representation through combination of a diagram encoder and a sequence encoder; designing a self-supervised pre-training task, and pre-training the semantic representation based on the training corpus; constructing a multi-language pre-training model based on the structure-improved recurrent neural tensor network and the multi-language embedding matrix; when a user inputs a natural language, generating a target language code by using the multi-language pre-training model; and target language code correction is carried out through conventional function testing and grammar checking. According to the method, multi-channel recursive combination and a hierarchical recursive expansion mechanism are combined with self-supervised pre-training, so that accurate generation and performability improvement of cross-language codes are realized.

The invention provides an anti-interference target detection method and system based on automatic driving scene multi-modal fusion, and the method comprises the steps: employing a semantic segmentation and instance segmentation combination method based on a camera image in an image view coding layer, and employing a Mask-RCNN model, and precisely judging a possible shielded object and region; in a feature processing link, a time self-attention mechanism is introduced, a feature map is weighted from a time dimension, an occluded object is focused, irrelevant information is inhibited, and the capability of capturing features of the occluded object is enhanced; besides, a caching mechanism is arranged in the model, a target confidence coefficient change method is adopted, the detection confidence coefficient of continuous video frames is stored, and the shielding condition is recognized by analyzing the confidence coefficient change condition of a target object in the continuous video frames. According to the invention, the detection precision and reliability of the shielded object are improved, and a more efficient and more accurate solution is provided for the object detection of the unmanned driving technology.

The invention provides a single Beidou differential positioning enhancement method based on NLOS signal identification, and the method comprises the steps: obtaining satellite signal data, and constructing a sky satellite map; a multipath signal identification model based on a Siamese neural network is constructed; unsupervised pre-training is carried out on a graph Transform encoder by adopting an asymmetrically enhanced sky satellite graph; performing feature extraction on the sky satellite feature map by using a trained map Transform encoder; constructing an ambiguity fixed strategy, and optimizing the ambiguity fixed strategy based on an NLOS recognition result; and calculating the position of the user by using the optimized ambiguity fixing strategy. According to the invention, the ambiguity fixing efficiency and reliability are further improved; and the position of the user is calculated through an optimized ambiguity fixing strategy, so that the positioning precision of the single Beidou satellite in the urban complex environment is improved.

The invention provides an enhanced generation method based on topology perception graph coding and adaptive sub-graph retrieval, which comprises the steps of multi-granularity semantic perception segmentation and knowledge graph construction, structural feature matrix construction and explicit topology position code generation, and constrained sub-graph diffusion and dynamic pruning based on semantic-topology joint scoring. Generating a graph-text consistency loss constraint based on attention matrix structured alignment; explicit topological position coding is adopted in a coding layer to avoid overhead and excessive smoothness caused by online GNN aggregation; in a retrieval layer, a connected sub-graph instead of a fragment node is used as an enhanced context; in an alignment layer, an internal attention matrix of a large language model is directly constrained to be consistent with a sub-graph adjacent matrix, and structured guidance is realized from a reasoning mechanism level, so that logic illusion is inhibited, and the multi-hop reasoning accuracy is improved.

The invention relates to the technical field of network security, and discloses an Internet of Things intrusion detection system and method based on a graph neural network. According to the system and the method, original flow data can be effectively collected from a network through a data collection and preprocessing module, IP address randomization, feature normalization and target coding are carried out, it is ensured that generated data is suitable for graph structure modeling, a graph structure construction module converts preprocessed data into graph structure representation, and the graph structure representation is realized. A clear network topology is formed by utilizing feature initialization of nodes and edges, the graph coding and feature extraction module deeply extracts and updates the features of the nodes and the edges through an improved graph encoder, the feature representation capability of the model is enhanced by introducing a contrast loss function, the recognition precision of potential intrusion behaviors is improved, and the recognition efficiency of the potential intrusion behaviors is improved. And the model training and classification judgment module performs accurate classification on the network traffic by using the trained model, identifies and outputs an intrusion detection result, and improves the accuracy and efficiency of intrusion detection in the Internet of Things environment.

The invention discloses an abnormal traffic flow detection method and system based on a dynamic graph, and is applied to the field of abnormal traffic flow detection. The method comprises the following steps: dividing a city into a plurality of sub-regions through city road network data, and obtaining vehicle trajectory data to construct a traffic flow tensor; constructing an approximate graph based on the traffic flow tensor and performing data enhancement to generate a local view and a global view of the traffic flow tensor; designing a comparative learning model with a time graph encoder to perform joint training on the multi-view data, and capturing space-time dependency features; and acquiring a real-time traffic flow embedding vector by adopting a sliding window, identifying an abnormal traffic mode through a local abnormal factor algorithm, and finally realizing the positioning of an urban abnormal traffic area. According to the method, through fusion of the time graph encoder and comparative learning, the accuracy of traffic anomaly detection in a complex road network environment is effectively improved, and reliable technical support is provided for urban traffic intelligent management.

The invention discloses a multi-lane heterogeneous traffic joint modeling method and system based on a GNN-PINN coupling architecture. The method comprises the following steps: constructing a space-time dynamic graph coding lane changing rule through a GNN, embedding a traffic flow physical constraint through a PINN, and solving a source item; according to the method, CAV behaviors are described by using improved IDM in the mid-microscopic view, an LWR equation subjected to GNN lane change flow correction is solved by using PINN in the macro view, GNN and PINN closed-loop optimization is realized by combining a double-path back propagation mechanism, and a model is optimized based on a residual loss function fusing physical loss, data driving items and boundary constraints. The system comprises a dynamic graph topology module, a GNN coding module, a PINN solving module and the like, the multi-lane traffic dynamic description precision can be improved, the multi-scale characteristic is considered, real-time simulation is supported, and the system is suitable for heterogeneous traffic modeling.

The invention discloses an online education cognition diagnosis method based on heterogeneous conceptual graph construction and modeling enhancement, and the method comprises the following steps: extracting practice records of a plurality of students, and extracting a plurality of knowledge concepts in the practice records; selecting knowledge concept pairs of which the similarity is higher than a threshold value, identifying the relation between knowledge concepts, and constructing an accurate heterogeneous conceptual graph through triple generation and triple evaluation; dividing the heterogeneous concept graph into three sub-graphs according to the relationship type; basic representations of the three sub-graphs are obtained through a sub-graph encoder; the important sub-graph features are dynamically identified through a self-adaptive selector; updating knowledge concepts and practice features through a graphic information aggregator; a diagnostic prediction of the student's knowledge grasp level is generated based on the response records and the graphical information. According to the method, the complex relation between knowledge concepts is mined, the accurate inference on the knowledge mastering level is improved, the basic sub-graph features are adaptively recognized, and graph representation is integrated.

The invention discloses a structure entropy model for detecting a social robot in real time, particularly relates to the technical field of big data mining, and is used for solving the problem that an existing social robot detection method is difficult to balance between detection precision and processing efficiency. User data is processed through a feature selection weighting module to screen key features and distribute weights, a feature stability evaluation module analyzes the contribution degree of the key features to a community structure to screen features with the most distinguishing power, and a multi-relation graph construction module constructs a user similarity graph based on the key features and the weights thereof. The coding tree community division module initializes a coding tree on the user similarity graph and generates a hierarchical community division structure through iterative calculation of structure entropy change; and the community level classification module forms a comprehensive score for each community fusion structure entropy and behavior statistical characteristics and judges a social robot community through comparison with an optimization threshold.

The invention belongs to the technical field of image data processing, and particularly relates to a wood surface defect detection method based on multi-view coding and a feature memory bank. The method comprises the steps that an original wood data set is acquired, an original wood image in the original wood data set is preprocessed, and the original wood image is converted into a Lab image; a wood surface defect detection model is constructed to process the Lab image, and a normal feature core set is obtained; after an anomaly detection framework is constructed and a feature core set is obtained, inputting a to-be-detected wood surface image into a wood surface defect detection model to extract local features, and calculating a nearest neighbor distance between each local feature and features in the normal feature core set; and generating an abnormal heat map of the to-be-detected image based on the nearest neighbor distance so as to realize defect detection and pixel-level positioning. The problems that the wood surface defect types are diversified, abnormal samples are difficult to cover, and the marking cost is high are solved.

The invention discloses a sub-graph reasoning method fusing logic rule learning and attack semantic enhancement. The method comprises the steps that an input layer dynamically integrates knowledge graph topology and an AMIE rule base, an initial k-hop sub-graph is generated, and structured input is provided for attack chain mining; the sub-graph extraction module is used for executing double confidence filtering, screening high-value attack chains and applying dictionary filtering to enhance semantic reliability; the sub-graph coding module adopts an entity perception update layer and a relationship aggregation evolution layer of a dual-channel mechanism to collaboratively model the spatial-temporal characteristics of an attack chain; the relation reasoning optimization module is used for dynamically injecting high confidence rules and optimizing triple scores; and the training optimization module is used for implementing task perception negative sampling. According to the subgraph reasoning method fusing logic rule learning and attack semantic enhancement, based on inductive reasoning and semantic perception modeling, by taking subgraph modeling guided by a logic path as a core, attack chain rules with high confidence in a training graph are mined, and the understanding ability of a model structure is improved.

This invention discloses a multimodal traffic flow prediction method based on multi-source data feature fusion, comprising: acquiring traffic data; dividing the traffic data into training and testing sets; constructing a traffic prediction model and training the traffic prediction model using the training set; and validating the traffic prediction model using the testing set. The traffic prediction model includes: a cloud map encoder for extracting features from cloud map data to obtain cloud map features; a spatiotemporal encoder for extracting features from spatiotemporal data to obtain spatiotemporal features; a fusion module for fusing cloud map features and spatiotemporal features to obtain fused features; and adding the cloud map features, spatiotemporal features, and fused features, inputting the sum to the spatiotemporal decoder for prediction to obtain the final prediction result. This invention integrates multi-dimensional and multi-faceted information, simplifies redundant information in the spatiotemporal feature extraction process, and achieves efficient and high-precision prediction, which can be widely applied in the field of traffic flow prediction technology.