Method and device for generating multi-scene time series data under zero-shot condition and computer equipment

Abstract

The application relates to a multi-scene time series data generation method, device and computer equipment under a zero sample condition. The method comprises the following steps: acquiring a city knowledge graph, a satellite image and scene description text, and performing encoding processing on the city knowledge graph to obtain city structure semantic information; generating city heterogeneous semantic information through a multi-modal encoding model, and performing alignment processing on the city structure semantic information and the city heterogeneous semantic information to obtain target structure semantic information and target heterogeneous semantic information; generating cross-domain spatio-temporal data through a conditional diffusion generation model based on the target structure semantic information and the target heterogeneous semantic information, and adjusting the cross-domain spatio-temporal data through a consistency regularization strategy to obtain multi-scene time series data. The method can improve the accuracy of time series data generation under a multi-scene condition of a zero sample.

Description

Technical Field

[0001] This application relates to the field of data mining technology, and in particular to a method, apparatus and computer device for generating multi-scenario time-series data under zero-sample conditions. Background Technology

[0002] Spatiotemporal data is widely used in applications such as urban computing, traffic analysis, and geographic information systems (GIS), including traffic flow data, shared bicycle data, crime data, and urban service demand data. However, due to high sensor deployment costs and limited data acquisition conditions, spatial observations are often sparse in practical applications. To address this issue, researchers have proposed spatiotemporal data generation techniques. These techniques aim to learn a conditional generation model based on observation data from a subset of spatial nodes, thereby generating spatiotemporal data for unobserved areas. However, this technique requires the model to generate reasonable spatiotemporal data for areas not yet observed. Therefore, improving the accuracy of the model's generated spatiotemporal data is a current research focus.

[0003] Traditional time series data generation methods are typically trained independently on data from a single domain. These methods rely on a large amount of labeled or fully observed data in the target domain. However, in practical applications, time series data generation tasks often need to be performed in situations where supervised data is lacking in the target region, i.e., generating data in zero-sample or weakly supervised scenarios. This results in low accuracy in generating time series data under multiple scenarios with zero samples. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for generating multi-scenario time-series data under zero-sample conditions to address the aforementioned technical problems.

[0005] Firstly, this application provides a method for generating multi-scenario time-series data under zero-sample conditions, including:

[0006] The city knowledge graph, satellite images, and scene description text are acquired, and the city knowledge graph is encoded to obtain the city structure semantic information.

[0007] Based on the satellite image and the scene description text, urban heterogeneous semantic information is generated through a multimodal coding model, and the urban structural semantic information and the urban heterogeneous semantic information are aligned to obtain target structural semantic information and target heterogeneous semantic information.

[0008] Based on the target structural semantic information and the target heterogeneous semantic information, cross-domain spatiotemporal data is generated through a conditional diffusion generation model. The cross-domain spatiotemporal data is then adjusted and processed through a consistency regularization strategy to obtain multi-scenario time-series data.

[0009] Optionally, the encoding process of the city knowledge graph to obtain city structural semantic information includes:

[0010] The city knowledge graph is modeled to obtain a set of city triples, and the city triples are initialized to obtain a city embedding matrix.

[0011] Based on the city embedding matrix, weighted encoding is performed using a graph attention network to obtain the city structure semantic information.

[0012] Optionally, the step of generating heterogeneous urban semantic information based on the satellite image and the scene description text through a multimodal coding model includes:

[0013] The satellite image and the scene description text are mapped into the same embedding space to obtain a standard satellite image and a standard scene description text;

[0014] The satellite image is divided into individual tile images, and each tile image is encoded into a visual feature vector using a visual encoder.

[0015] The visual feature vectors are semantically embedded to obtain semantic embedding information corresponding to each tile image, and the sub-descriptive texts related to each domain in the scene description text are encoded to obtain domain embedding information corresponding to each domain.

[0016] The semantic embedding information corresponding to each tile image and the domain embedding information corresponding to each domain are used as urban heterogeneous semantic information.

[0017] Optionally, the alignment processing of the urban structural semantic information and the urban heterogeneous semantic information to obtain target structural semantic information and target heterogeneous semantic information includes:

[0018] The urban structural semantic information and the urban heterogeneous semantic information are divided into regions to obtain the sub-urban structural semantic information and the sub-urban heterogeneous semantic information corresponding to each region.

[0019] For each region, a semantic reference space for the heterogeneous semantic information of the sub-cities corresponding to the region is identified, and the structural semantic information of the sub-cities is aligned to the semantic reference space through a spatial alignment model to obtain the structural semantic information of the sub-targets.

[0020] The sub-target structural semantic information of all regions is used as the target structural semantic information, and the heterogeneous semantic information of cities is directly used as the target heterogeneous semantic information.

[0021] Optionally, the step of generating cross-domain spatiotemporal data based on the target structural semantic information and the target heterogeneous semantic information through a conditional diffusion generation model includes:

[0022] By using a unified temporal coding model, temporal coding is performed on the target structural semantic information and the heterogeneous semantic information to obtain latent structural representation information and latent heterogeneous representation information.

[0023] The latent structural representation information and the latent heterogeneous representation information are mapped back to the original time domain using a time decoder to obtain the structural reconstruction sequence and the heterogeneous reconstruction sequence.

[0024] Based on the structural reconstruction sequence and the heterogeneous reconstruction sequence, cross-domain spatiotemporal data are obtained by guiding trajectory generation through a conditional diffusion model.

[0025] Optionally, the cross-domain spatiotemporal data is adjusted using a consistency regularization strategy to obtain multi-scenario time-series data, including:

[0026] Based on the aforementioned cross-domain spatiotemporal data, perturbation procedures are used to construct perturbed conditional embedded spatiotemporal data.

[0027] Based on the cross-domain spatiotemporal data and the conditionally embedded spatiotemporal data, a noise evaluation procedure is used to generate a first noise output result corresponding to the cross-domain spatiotemporal data and a second noise output result corresponding to the conditionally embedded spatiotemporal data.

[0028] Based on the first noise output result and the second noise output result, the regularized loss data of the cross-domain spatiotemporal data is calculated, and the cross-domain spatiotemporal data is adjusted based on the regularized loss data to obtain multi-scene time series data.

[0029] Secondly, this application also provides a multi-scenario time-series data generation device under zero-sample conditions, comprising:

[0030] The acquisition module is used to acquire urban knowledge graphs, satellite images, and scene description text, and to encode the urban knowledge graphs to obtain urban structural semantic information.

[0031] The alignment module is used to generate urban heterogeneous semantic information based on the satellite image and the scene description text through a multimodal coding model, and to align the urban structural semantic information and the urban heterogeneous semantic information to obtain target structural semantic information and target heterogeneous semantic information.

[0032] The adjustment module is used to generate cross-domain spatiotemporal data based on the target structural semantic information and the target heterogeneous semantic information through a conditional diffusion generation model, and to adjust the cross-domain spatiotemporal data through a consistency regularization strategy to obtain multi-scenario time series data.

[0033] Optionally, the acquisition module is specifically used for:

[0034] The city knowledge graph is modeled to obtain a set of city triples, and the city triples are initialized to obtain a city embedding matrix.

[0035] Based on the city embedding matrix, weighted encoding is performed using a graph attention network to obtain the city structure semantic information.

[0036] Optionally, the alignment module is specifically used for:

[0037] The satellite image and the scene description text are mapped into the same embedding space to obtain a standard satellite image and a standard scene description text;

[0038] The satellite image is divided into individual tile images, and each tile image is encoded into a visual feature vector using a visual encoder.

[0039] The visual feature vectors are semantically embedded to obtain semantic embedding information corresponding to each tile image, and the sub-descriptive texts related to each domain in the scene description text are encoded to obtain domain embedding information corresponding to each domain.

[0040] The semantic embedding information corresponding to each tile image and the domain embedding information corresponding to each domain are used as urban heterogeneous semantic information.

[0041] Optionally, the alignment module is specifically used for:

[0042] The urban structural semantic information and the urban heterogeneous semantic information are divided into regions to obtain the sub-urban structural semantic information and the sub-urban heterogeneous semantic information corresponding to each region.

[0043] For each region, a semantic reference space for the heterogeneous semantic information of the sub-cities corresponding to the region is identified, and the structural semantic information of the sub-cities is aligned to the semantic reference space through a spatial alignment model to obtain the structural semantic information of the sub-targets.

[0044] The sub-target structural semantic information of all regions is used as the target structural semantic information, and the heterogeneous semantic information of cities is directly used as the target heterogeneous semantic information.

[0045] Optionally, the adjustment module is specifically used for:

[0046] By using a unified temporal coding model, temporal coding is performed on the target structural semantic information and the heterogeneous semantic information to obtain latent structural representation information and latent heterogeneous representation information.

[0047] The latent structural representation information and the latent heterogeneous representation information are mapped back to the original time domain using a time decoder to obtain the structural reconstruction sequence and the heterogeneous reconstruction sequence.

[0048] Based on the structural reconstruction sequence and the heterogeneous reconstruction sequence, cross-domain spatiotemporal data are obtained by guiding trajectory generation through a conditional diffusion model.

[0049] Optionally, the adjustment module is specifically used for:

[0050] Based on the aforementioned cross-domain spatiotemporal data, perturbation procedures are used to construct perturbed conditional embedded spatiotemporal data.

[0051] Based on the cross-domain spatiotemporal data and the conditionally embedded spatiotemporal data, a noise evaluation procedure is used to generate a first noise output result corresponding to the cross-domain spatiotemporal data and a second noise output result corresponding to the conditionally embedded spatiotemporal data.

[0052] Based on the first noise output result and the second noise output result, the regularized loss data of the cross-domain spatiotemporal data is calculated, and the cross-domain spatiotemporal data is adjusted based on the regularized loss data to obtain multi-scene time series data.

[0053] Thirdly, this application provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described in any one of the first aspects.

[0054] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0055] Fifthly, this application provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0056] The aforementioned method, apparatus, and computer equipment for generating multi-scenario time-series data under zero-sample conditions acquire urban knowledge graphs, satellite images, and scene description texts, and encode the urban knowledge graphs to obtain urban structural semantic information. Based on the satellite images and scene description texts, a multimodal coding model is used to generate urban heterogeneous semantic information, and the urban structural semantic information and the urban heterogeneous semantic information are aligned to obtain target structural semantic information and target heterogeneous semantic information. Based on the target structural semantic information and the target heterogeneous semantic information, a conditional diffusion generation model is used to generate cross-domain spatiotemporal data, and a consistency regularization strategy is used to adjust the cross-domain spatiotemporal data to obtain multi-scenario time-series data. This solution can comprehensively generate multi-scenario time-series data in multi-domain urban spatiotemporal data scenarios by acquiring urban knowledge graphs, satellite images, and scene description texts. This avoids the problem that traditional time-series data generation tasks often rely on a large amount of labeled or fully observed data in the target domain, thus achieving zero-shot generation of time-series data under multi-scenario conditions. Secondly, when generating time-series data, this solution encodes urban structural semantic information and urban heterogeneous semantic information from urban knowledge graphs, satellite images, and scene description texts respectively, and performs alignment processing to achieve a unified representation of different modal information. This ensures that the different modal information acquired can be used uniformly when generating time-series data later. Finally, this solution uses a conditional diffusion generation model to diffuse and generate the aforementioned urban structural semantic information and urban heterogeneous semantic information, thereby enabling pre-training on multiple domain data and learning transferable spatiotemporal patterns, thus supporting cross-regional and cross-domain spatiotemporal data generation. Finally, this solution employs a consistency regularization strategy to ensure the continuity of the generated results under minor perturbations of cross-domain spatiotemporal data. This results in continuous, smooth, and consistent cross-domain time series data, improving the stability and authenticity of the generated results. Consequently, it comprehensively enhances the accuracy of time series data generation under zero-sample, multi-scenario conditions. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating a method for generating multi-scenario time-series data under zero-sample conditions in one embodiment.

[0059] Figure 2A flowchart illustrating an example of generating time-series data across multiple scenarios under zero-sample conditions in one embodiment;

[0060] Figure 3 This is a structural block diagram of a multi-scenario time-series data generation device under zero-sample conditions in one embodiment.

[0061] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] The zero-sample-condition multi-scenario time-series data generation method provided in this application can be applied to a zero-sample-condition multi-scenario time-series data generation system. This system can be applied to a terminal, which can be, but is not limited to, various personal computers, laptops, mid-range computers, etc. The terminal can generate multi-scenario time-series data by comprehensively utilizing urban knowledge graphs, satellite images, and scene description texts in multi-domain urban spatiotemporal data scenarios. This avoids the problem that traditional time-series data generation tasks often rely on a large amount of labeled or fully observed data in the target domain, thus achieving zero-sample generation of time-series data under multi-scenario conditions. Secondly, when generating time-series data, this solution encodes urban knowledge graphs, satellite images, and scene description texts to obtain urban structural semantic information and urban heterogeneous semantic information, and performs alignment processing to achieve a unified representation of different modal information. This ensures that the different modal information obtained can be uniformly utilized when generating time-series data later. Finally, this solution uses a conditional diffusion generation model to diffuse and generate the aforementioned urban structural semantic information and urban heterogeneous semantic information, thereby enabling pre-training on multiple domain data and learning transferable spatiotemporal patterns, thus supporting cross-regional and cross-domain spatiotemporal data generation. Finally, this solution employs a consistency regularization strategy to ensure the continuity of the generated results under minor perturbations of cross-domain spatiotemporal data. This results in continuous, smooth, and consistent cross-domain time series data, improving the stability and authenticity of the generated results. Consequently, it comprehensively enhances the accuracy of time series data generation under zero-sample, multi-scenario conditions.

[0064] In one exemplary embodiment, such as Figure 1 As shown, a method for generating multi-scenario time-series data under zero-sample conditions is provided. Taking the application of this method to a terminal as an example, the method includes the following steps S101 to S103. Wherein:

[0065] Step S101: Obtain the city knowledge graph, satellite images, and scene description text, and encode the city knowledge graph to obtain the city structure semantic information.

[0066] In this embodiment, the terminal responds to the information input operation of the staff, acquires a structured urban knowledge graph including different regions and fields, receives high-resolution satellite images transmitted in real time, and then acquires descriptive domain text manually input by the staff for each region and field, obtaining scene description text. Next, the terminal uses a graph attention network to collect the urban context semantic information associated with nodes in the urban knowledge graph, and then encodes it to obtain urban structural semantic information. This urban knowledge graph is a multi-relationship graph constructed following the UUKG framework. The specific encoding process will be explained in detail later.

[0067] Step S102: Based on satellite images and scene description text, urban heterogeneous semantic information is generated through a multimodal coding model, and the urban structural semantic information and urban heterogeneous semantic information are aligned to obtain target structural semantic information and target heterogeneous semantic information.

[0068] In this embodiment, the terminal generates heterogeneous urban semantic information based on satellite imagery and scene description text using a multimodal coding model. It then aligns the urban structural semantic information and the heterogeneous urban semantic information to obtain target structural semantic information and target heterogeneous semantic information. The multimodal coding model is a Multimodal Large Language Model (MLLM). The terminal first extracts multimodal representations from satellite imagery and domain-level text descriptions. Then, it uses an MLLM to project these conditional inputs into a unified embedding space to obtain the heterogeneous urban semantic information. This heterogeneous urban semantic information consists of semantic embedding vectors representing different regions and domains, including semantic embedding information corresponding to different tile images and domain embedding information corresponding to different regions and domains (i.e., domains). The alignment process maps the urban structural semantic information uniformly to the semantic space containing the heterogeneous urban semantic information. Targeted heterogeneous semantic information is generated by combining image modalities, while urban structural semantic information is generated from structured data modalities corresponding to knowledge graphs. Mapping the semantic information of the structured data modalities to the semantic space of the image modalities can effectively reduce mapping costs and improve mapping efficiency while ensuring alignment processing. The specific mapping process will be explained in detail later.

[0069] Step S103: Based on the target structural semantic information and the target heterogeneous semantic information, cross-domain spatiotemporal data is generated through a conditional diffusion generation model, and the cross-domain spatiotemporal data is adjusted and processed through a consistency regularization strategy to obtain multi-scenario time series data.

[0070] In this embodiment, the terminal generates cross-domain spatiotemporal data based on target structural semantic information and target heterogeneous semantic information through a conditional diffusion generation model. A consistency regularization strategy is then used to adjust and process the cross-domain spatiotemporal data to obtain multi-scene time-series data. The conditional diffusion generation model includes a unified temporal coding model and a conditional diffusion model. The target structural semantic information includes sub-target structural semantic information for each region, and the target heterogeneous semantic information includes sub-target heterogeneous semantic information for different regions. The unified temporal coding model uses a sliding window to convert the original sequences of sub-target structural semantic information and sub-target heterogeneous semantic information from different regions into a fixed-scale latent representation, enabling a shared diffusion backbone network applicable to models across all domains. The conditional diffusion model uses an Elicuated Diffusion Model (EDM) as its backbone network, introducing a conditional mean to guide the generation of trajectories for different regions and domain scenarios, thus forming a model of cross-domain spatiotemporal data. The specific application process of the model will be explained in detail later. The regularization strategy involves introducing small perturbations into the cross-domain spatiotemporal data to generate noisy data. Then, based on this noisy data and the cross-domain spatiotemporal data, the data values ​​that need to be adjusted for consistency are determined so that the adjusted multi-scenario time-series data is consistent in distribution, continuity, and smoothness. The specific generation process will be explained in detail later.

[0071] Based on the above scheme, multi-scenario time-series data can be generated comprehensively from urban knowledge graphs, satellite images, and scene description texts in multi-domain urban spatiotemporal data scenarios. This avoids the problem that traditional time-series data generation tasks often rely on a large amount of labeled or fully observed data in the target domain, thus achieving zero-shot generation of time-series data under multi-scenario conditions. Secondly, when generating time-series data, this scheme encodes urban knowledge graphs, satellite images, and scene description texts to obtain urban structural semantic information and urban heterogeneous semantic information, and performs alignment processing to achieve a unified representation of different modal information. This ensures that the different modal information obtained can be uniformly utilized when generating time-series data later. Finally, this scheme uses a conditional diffusion generation model to diffuse and generate the above-mentioned urban structural semantic information and urban heterogeneous semantic information, thereby enabling pre-training on multiple domain data and learning transferable spatiotemporal patterns, thus supporting cross-regional and cross-domain spatiotemporal data generation. Finally, this solution employs a consistency regularization strategy to ensure the continuity of the generated results under minor perturbations of cross-domain spatiotemporal data. This results in continuous, smooth, and consistent cross-domain time series data, improving the stability and authenticity of the generated results. Consequently, it comprehensively enhances the accuracy of time series data generation under zero-sample, multi-scenario conditions.

[0072] Optionally, the city knowledge graph is encoded to obtain city structural semantic information, including: modeling the city knowledge graph to obtain a set of city triples, initializing the city triples to obtain a city embedding matrix; and performing weighted encoding processing on the city embedding matrix through a graph attention network to obtain city structural semantic information.

[0073] In this embodiment, the terminal models the city knowledge graph to obtain a set of city triples, and initializes these city triples to obtain a city embedding matrix. Specifically, the formal definition of this city knowledge graph is:

[0074]

[0075]

[0076] in, , and Each triple represents a set of entity, relation, and fact triples. Represents the head entity There is a relationship between the tail entity t. This knowledge graph encodes the environmental context of a city in detail, including hierarchical spatial connectivity (such as administrative division affiliation) and functionally important points of interest (POIs), which are crucial for characterizing the spatiotemporal generation patterns of different regions.

[0077] The above modeling process involves the terminal processing all entities... Initialize as an entity embedding matrix and all relationships Initialize as a relation embedding matrix . and These represent the number of entities and relations, respectively. It is the dimension of the conditional embedding space.

[0078] Finally, based on the city embedding matrix, the terminal performs weighted encoding processing using a graph attention network to obtain the semantic information of the city structure. Specifically, for the central entity... and its neighbors Each neighbor contributes a relationship-aware message, defined as follows: ,in and Representing the relationship Tail-end entity Embedded entity. The final embedding is obtained by aggregating all its weighted neighbor messages, calculated as follows:

[0079]

[0080] in, It is an activation function. It is a triplet The message weight is calculated using an attention mechanism:

[0081]

[0082] Attention scoring function Defined as:

[0083]

[0084] in, and It is a trainable projection matrix. It is a learnable attention vector. This is the dimension of the hidden layer embedding. In this way, for each region entity... It generates embedding representations rich in structural semantics. Finally, the terminal will use all the embedding representations of each region as a city embedding matrix.

[0085] Based on the above scheme, a city embedding matrix is ​​constructed by building triples and using the graph attention network method adopted in this scheme, thereby improving the accuracy and comprehensiveness of extracting the city context semantics associated with nodes in the city knowledge graph.

[0086] Optionally, based on satellite imagery and scene description text, a multimodal coding model is used to generate heterogeneous semantic information for the city. This includes: mapping satellite imagery and scene description text to the same embedding space to obtain standard satellite imagery and standard scene description text; splitting the satellite imagery into individual tile images and encoding each tile image into a visual feature vector using a visual encoder; performing semantic embedding processing on each visual feature vector to obtain semantic embedding information corresponding to each tile image; encoding sub-descriptive text related to each domain in the scene description text to obtain domain embedding information corresponding to each domain; and using the semantic embedding information corresponding to each tile image and the domain embedding information corresponding to each domain as heterogeneous semantic information for the city.

[0087] In this embodiment, the terminal maps satellite images and scene description text into the same embedding space to obtain standard satellite images and standard scene description text. Since the satellite images are provided in the form of georeferenced image tiles, they are essentially images composed of multiple "tile" images. Therefore, the terminal splits the satellite image into individual tile images. The set formed by each tile image is... .

[0088] Then, the terminal encodes each tile image into a visual feature vector using a visual encoder. Specifically, the terminal encodes the tile images into visual feature vectors using a pre-trained CLIP model's visual encoder. .

[0089] Next, the terminal performs semantic embedding processing on each visual feature vector to obtain semantic embedding information corresponding to each tile image, and encodes the domain-related sub-descriptive text in the scene description text to obtain domain embedding information corresponding to each domain. Specifically, the terminal projects the data into the language embedding space using an MLLM multimodal projector, and then, through forward propagation of the language model, obtains tile-level semantic embeddings:

[0090]

[0091] in, A visual-speech projector representing MLLM. It represents the language model itself.

[0092] To align the tile embeddings with other conditional modes, this invention introduces a lightweight MLLM adapter. The satellite image embedding of a region is obtained by average pooling the embeddings of all tiles within that region:

[0093]

[0094] in, It consists of a linear projection layer and a subsequent layer normalization operation. express Normalization.

[0095] To characterize the domain-specific semantics of each dataset, the terminal uses the same MLLM to encode the descriptive text associated with each domain. Given a piece of domain descriptive text... The terminal extracts the hidden states of the last layer of the language model and obtains the domain embeddings (i.e., the domain embedding information corresponding to each domain) through average pooling.

[0096]

[0097] in, This indicates the number of tokens in the text. Indicates the first The last hidden state corresponding to each token. The resulting embedding. As a global condition, it describes the overall characteristics of this spacetime domain.

[0098] Finally, the terminal uses the semantic embedding information corresponding to each tile image and the domain embedding information corresponding to each domain as the heterogeneous semantic information of the city.

[0099] Based on the above scheme, satellite images and scene description texts are encoded using a multimodal large language model to obtain domain embedding information that can represent scenes in different regions and fields, thereby improving the comprehensiveness of extracting more heterogeneous semantic information that goes beyond structured urban knowledge.

[0100] Optionally, the urban structural semantic information and urban heterogeneous semantic information are aligned to obtain target structural semantic information and target heterogeneous semantic information. This includes: dividing the urban structural semantic information and urban heterogeneous semantic information into regions to obtain sub-urban structural semantic information and sub-urban heterogeneous semantic information corresponding to each region; for each region, identifying the semantic reference space of the sub-urban heterogeneous semantic information corresponding to the region, and aligning the sub-urban structural semantic information to the semantic reference space through a spatial alignment model to obtain sub-target structural semantic information; using the sub-target structural semantic information of all regions as target structural semantic information, and directly using urban heterogeneous semantic information as target heterogeneous semantic information.

[0101] In this embodiment, the terminal performs regional segmentation of urban structural semantic information and urban heterogeneous semantic information to obtain sub-urban structural semantic information and sub-urban heterogeneous semantic information corresponding to each region. Then, for each region, the terminal identifies the semantic reference space of the sub-urban heterogeneous semantic information corresponding to the region, and aligns the sub-urban structural semantic information to the semantic reference space using a spatial alignment model to obtain the sub-target structural semantic information. Specifically, the terminal sets... The embedding of region entity n is obtained through the knowledge graph encoder. Embedded into the corresponding satellite imagery. For Each region, and the terminal defines a similarity matrix. ,in This is a temperature parameter. Positive sample pairs consist of embeddings from two modalities originating from the same region, while negative sample pairs consist of embeddings from different regions. The alignment loss function is defined as:

[0102]

[0103] in, It is a knowledge-based contrastive loss, calculated using the following formula:

[0104]

[0105] and A corresponding symmetric definition is adopted. Hyperparameters Used to control the anchoring direction of the shared semantic space. When When the value is small, the alignment process will be more biased towards the MLLM modality, that is, treating the multimodal representation (image and text) as a semantic reference space and narrowing the distance between the knowledge graph embedding and this space. This alignment objective jointly optimizes the GAT-based knowledge graph encoder and MLLM adapter, prompting the region embeddings from different modalities to converge in a shared conditional space, providing a unified, semantically aligned conditional input for subsequent diffusion models.

[0106] Finally, the terminal uses the sub-target structural semantic information of all regions as the target structural semantic information, and directly uses the heterogeneous semantic information of cities as the target heterogeneous semantic information.

[0107] Based on the above scheme, by aligning the city knowledge graph with the semantic information identified by satellite images into the same semantic space, the processing efficiency during subsequent data processing is ensured.

[0108] Optionally, based on the target structural semantic information and the target heterogeneous semantic information, cross-domain spatiotemporal data is generated through a conditional diffusion generation model, including: performing time encoding processing on the target structural semantic information and the heterogeneous semantic information through a unified time encoding model to obtain potential structural representation information and potential heterogeneous representation information; mapping the potential structural representation information and the potential heterogeneous representation information back to the original time domain through a time decoder to obtain structural reconstruction sequence and heterogeneous reconstruction sequence; and guiding trajectory generation through a conditional diffusion model based on the structural reconstruction sequence and the heterogeneous reconstruction sequence to obtain cross-domain spatiotemporal data.

[0109] In this embodiment, the terminal uses a unified temporal coding model to perform temporal coding on the target structural semantic information and heterogeneous semantic information to obtain potential structural representation information and potential heterogeneous representation information.

[0110] Specifically, the terminal is for the domain area in The observed spatiotemporal data are denoted as The length of time In different fields The differences may exist. The terminal uses a method defined by the window length. and step length Parameterized sliding window time expansion operator This operator transforms a one-dimensional time series into a two-dimensional representation:

[0111]

[0112] Each window segment is defined as follows:

[0113]

[0114] Number of windows This representation preserves local temporal continuity while providing a unified underlying structure for subsequent diffusion modeling. Due to differences in domains... The values ​​may differ; the terminal fills the expanded tensor to form a square representation. ,in This filled-in The input is then projected into a latent space as input to the diffusion process. To ensure architectural consistency, the diffusion model operates at a fixed internal resolution. When the input resolution differs, bilinear interpolation is applied before and after the denoising network without altering the temporal semantics.

[0115] Finally, the terminal uses a temporal decoder to map the latent structural representation information and the latent heterogeneous representation information back to the original time domain, respectively, to obtain the structural reconstruction sequence and the heterogeneous reconstruction sequence. Specifically, after the diffusion process is complete, a temporal decoder performs an inversion unrolling operation to map the denoised latent representation back to the original time domain, generating a reconstruction sequence with the original time length. .

[0116] Subsequently, based on the structural reconstruction sequence and the heterogeneous reconstruction sequence, the terminal guides trajectory generation through a conditional diffusion model to obtain cross-domain spatiotemporal data.

[0117] Specifically, the terminal introduces a conditional mean to guide the generation of trajectories. This design reflects the fact that spatiotemporal data often exhibits daily average patterns determined by city and domain semantics.

[0118] Given a unified multimodal conditional embedding , and The terminal first predicts a conditional latent prior mean. This mean represents the expected potential representation of the corresponding region and field under the daily average pattern. Let... This is a potential representation of a noise-free spatiotemporal sequence obtained from a time encoder. This invention does not assume that the data distribution is zero-centered, but rather defines a forward diffusion process with a mean shift:

[0119]

[0120] here This represents the noise level. Equivalently, the diffusion process is applied to the residual latent variables. The above formula can be transformed into This form preserves the standard EDM forward process while shifting the diffusion trajectory to a neighborhood around the conditionally induced latent prior mean. To ensure compatibility with the standard U-Net denoising architecture, It is used as input to the denoising network.

[0121] The denoising network is conditional on both centralized latent variables and multimodal context. Specifically, a unified conditional embedding is injected into each module of U-Net via an adaptive affine transformation. Following the parameterization method of EDM, the denoised output is reconstructed as follows:

[0122]

[0123] in, This represents a conditional denoising network. and This is a scaling factor defined in EDM that depends on the noise level. The predicted mean. It is explicitly added back to recover the final latent representation. Following the training paradigm of EDM, the optimization objective of the diffusion model is to reconstruct a clean latent representation from the noisy input under different noise levels. Specifically, this invention uses a predefined distribution... Medium sampling noise level The denoising network then generates noisy latent variables based on the mean shift process. The denoising network predicts clean latent variables using the formula described above. The training objective is defined as the weighted mean squared error:

[0124]

[0125] in, It is a weighted function dependent on the noise level, used to balance contributions from different noise scales. The diffusion model is first pre-trained on the source domain and then fine-tuned on the target domain using the same conditional diffusion objective. This shared formulaic approach enables cross-domain knowledge transfer while adapting the generative prior to the spatiotemporal characteristics of a specific region.

[0126] Based on the above scheme, by adopting a conditional diffusion framework to model spatiotemporal data across urban domains, it is possible to generate spatiotemporal characteristics adapted to specific regions while transferring knowledge across domains, thereby improving the accuracy of spatiotemporal data generation for multiple scenarios.

[0127] Optionally, a consistency regularization strategy is used to adjust and process the cross-domain spatiotemporal data to obtain multi-scenario time-series data, including: constructing perturbed conditional embedding spatiotemporal data based on the cross-domain spatiotemporal data through a noise perturbation procedure; generating a first noise output result corresponding to the cross-domain spatiotemporal data and a second noise output result of the conditional embedding spatiotemporal data based on the cross-domain spatiotemporal data and the conditional embedding spatiotemporal data through a noise evaluation procedure; calculating the regularization loss data of the cross-domain spatiotemporal data based on the first noise output result and the second noise output result, and adjusting the cross-domain spatiotemporal data based on the regularization loss data to obtain multi-scenario time-series data.

[0128] In this embodiment, the terminal constructs perturbed conditional embedding spatiotemporal data based on cross-domain spatiotemporal data through a noise perturbation procedure. Then, based on the cross-domain spatiotemporal data and the conditional embedding spatiotemporal data, the terminal generates a first noise output result corresponding to the cross-domain spatiotemporal data and a second noise output result for the conditional embedding spatiotemporal data through a noise evaluation procedure. Specifically, let... , Let e^(TXT) ∈ ℝ^(d_cond) be a uniform modal embedding. The terminal constructs a perturbed conditional embedding by injecting small amounts of Gaussian noise:

[0129]

[0130] In the above formula, , Control the amplitude of the perturbation. This operation samples neighboring points in a continuous conditional space without altering the semantic structure of the conditional information. For a given noise level... The terminal is under the same noisy potential input The outputs of the denoising network under the original conditions and the perturbation conditions are evaluated separately:

[0131]

[0132]

[0133] The difference between these two outputs strips away random fluctuations caused by input noise and specifically isolates the effects of conditional perturbations.

[0134] Finally, based on the first noise output result and the second noise output result, the terminal calculates the regularized loss data of the cross-domain spatiotemporal data, and adjusts the cross-domain spatiotemporal data based on the regularized loss data to obtain multi-scenario time-series data. The formula for the above calculation is as follows:

[0135]

[0136] This loss function penalizes the denoised output's oversensitivity to small perturbations in the conditional space. The final training objective combines the standard EDM prediction loss with the proposed consistency regularization term:

[0137]

[0138] in The strength of consistency constraints is controlled. The terminal adds the above consistency regularization terms to the cross-domain spatiotemporal data to obtain multi-scenario time-series data.

[0139] Based on the above scheme, the continuity of the generated results under slight perturbations of cross-domain spatiotemporal data is achieved through a consistency regularization strategy. This makes the generated cross-domain time series data continuous, smooth, and consistent, thereby improving the stability and authenticity of the generated results and comprehensively enhancing the accuracy of time series data generation under zero-sample multi-scenario conditions.

[0140] This application also provides an example of generating multi-scenario time-series data under zero-sample conditions, such as... Figure 2 As shown, the specific processing procedure includes the following steps:

[0141] Step S201: Obtain the city knowledge graph, satellite images, and scene description text.

[0142] Step S202: Model the city knowledge graph to obtain a set of city triples, and initialize the city triples to obtain the city embedding matrix.

[0143] Step S203: Based on the city embedding matrix, weighted encoding is performed using a graph attention network to obtain the city structure semantic information.

[0144] Step S204: Map the satellite image and the scene description text into the same embedding space to obtain a standard satellite image and a standard scene description text.

[0145] Step S205: The satellite image is split into individual tile images, and each tile image is encoded into a visual feature vector using a visual encoder.

[0146] Step S206: Perform semantic embedding processing on each visual feature vector to obtain semantic embedding information corresponding to each tile image, and encode the sub-descriptive text related to each domain in the scene description text to obtain the domain embedding information corresponding to each domain.

[0147] Step S207: The semantic embedding information corresponding to each tile image and the domain embedding information corresponding to each domain are used as urban heterogeneous semantic information.

[0148] Step S208: The urban structural semantic information and the urban heterogeneous semantic information are divided into regions to obtain the sub-city structural semantic information and the sub-city heterogeneous semantic information corresponding to each region.

[0149] Step S209: For each region, identify the semantic reference space of the heterogeneous semantic information of the sub-cities corresponding to the region, and align the structural semantic information of the sub-cities to the semantic reference space through a spatial alignment model to obtain the structural semantic information of the sub-targets.

[0150] Step S210: Use the sub-target structural semantic information of all regions as target structural semantic information, and directly use the heterogeneous semantic information of cities as target heterogeneous semantic information.

[0151] Step S211: Using a unified temporal coding model, temporal coding is performed on the target structural semantic information and heterogeneous semantic information to obtain potential structural representation information and potential heterogeneous representation information.

[0152] Step S212: Using a time decoder, the latent structural representation information and the latent heterogeneous representation information are mapped back to the original time domain to obtain the structural reconstruction sequence and the heterogeneous reconstruction sequence.

[0153] Step S213: Based on the structural reconstruction sequence and the heterogeneous reconstruction sequence, trajectory generation is guided by the conditional diffusion model to obtain cross-domain spatiotemporal data.

[0154] Step S214: Based on cross-domain spatiotemporal data, construct perturbed conditional embedded spatiotemporal data through a noise perturbation procedure.

[0155] Step S215: Based on cross-domain spatiotemporal data and conditionally embedded spatiotemporal data, a noise evaluation procedure is used to generate a first noise output result corresponding to the cross-domain spatiotemporal data and a second noise output result corresponding to the conditionally embedded spatiotemporal data.

[0156] Step S216: Based on the first noise output result and the second noise output result, calculate the regularized loss data of the cross-domain spatiotemporal data, and adjust the cross-domain spatiotemporal data based on the regularized loss data to obtain multi-scenario time series data.

[0157] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0158] Based on the same inventive concept, this application also provides a zero-sample-condition multi-scene time-series data generation apparatus for implementing the zero-sample-condition multi-scene time-series data generation method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method. Therefore, the specific limitations in one or more zero-sample-condition multi-scene time-series data generation apparatus embodiments provided below can be found in the limitations of the zero-sample-condition multi-scene time-series data generation method described above, and will not be repeated here.

[0159] In one exemplary embodiment, such as Figure 3 As shown, a multi-scenario time-series data generation device under zero-sample conditions is provided, including: an acquisition module 310, an alignment module 320, and an adjustment module 330, wherein:

[0160] The acquisition module 310 is used to acquire urban knowledge graphs, satellite images, and scene description text, and to encode the urban knowledge graphs to obtain urban structural semantic information.

[0161] Alignment module 320 is used to generate urban heterogeneous semantic information based on the satellite image and the scene description text through a multimodal coding model, and to perform alignment processing on the urban structural semantic information and the urban heterogeneous semantic information to obtain target structural semantic information and target heterogeneous semantic information.

[0162] The adjustment module 330 is used to generate cross-domain spatiotemporal data based on the target structural semantic information and the target heterogeneous semantic information through a conditional diffusion generation model, and to adjust the cross-domain spatiotemporal data through a consistency regularization strategy to obtain multi-scenario time series data.

[0163] Optionally, the acquisition module 310 is specifically used for:

[0164] The city knowledge graph is modeled to obtain a set of city triples, and the city triples are initialized to obtain a city embedding matrix.

[0165] Based on the city embedding matrix, weighted encoding is performed using a graph attention network to obtain the city structure semantic information.

[0166] Optionally, the alignment module 320 is specifically used for:

[0167] The satellite image and the scene description text are mapped into the same embedding space to obtain a standard satellite image and a standard scene description text;

[0168] The satellite image is divided into individual tile images, and each tile image is encoded into a visual feature vector using a visual encoder.

[0169] The visual feature vectors are semantically embedded to obtain semantic embedding information corresponding to each tile image, and the sub-descriptive texts related to each domain in the scene description text are encoded to obtain domain embedding information corresponding to each domain.

[0170] The semantic embedding information corresponding to each tile image and the domain embedding information corresponding to each domain are used as urban heterogeneous semantic information.

[0171] Optionally, the alignment module 320 is specifically used for:

[0172] The urban structural semantic information and the urban heterogeneous semantic information are divided into regions to obtain the sub-urban structural semantic information and the sub-urban heterogeneous semantic information corresponding to each region.

[0173] For each region, a semantic reference space for the heterogeneous semantic information of the sub-cities corresponding to the region is identified, and the structural semantic information of the sub-cities is aligned to the semantic reference space through a spatial alignment model to obtain the structural semantic information of the sub-targets.

[0174] The sub-target structural semantic information of all regions is used as the target structural semantic information, and the heterogeneous semantic information of cities is directly used as the target heterogeneous semantic information.

[0175] Optionally, the adjustment module 330 is specifically used for:

[0176] By using a unified temporal coding model, temporal coding is performed on the target structural semantic information and the heterogeneous semantic information to obtain latent structural representation information and latent heterogeneous representation information.

[0177] The latent structural representation information and the latent heterogeneous representation information are mapped back to the original time domain using a time decoder to obtain the structural reconstruction sequence and the heterogeneous reconstruction sequence.

[0178] Based on the structural reconstruction sequence and the heterogeneous reconstruction sequence, cross-domain spatiotemporal data are obtained by guiding trajectory generation through a conditional diffusion model.

[0179] Optionally, the adjustment module 330 is specifically used for:

[0180] Based on the aforementioned cross-domain spatiotemporal data, perturbation procedures are used to construct perturbed conditional embedded spatiotemporal data.

[0181] Based on the cross-domain spatiotemporal data and the conditionally embedded spatiotemporal data, a noise evaluation procedure is used to generate a first noise output result corresponding to the cross-domain spatiotemporal data and a second noise output result corresponding to the conditionally embedded spatiotemporal data.

[0182] Based on the first noise output result and the second noise output result, the regularized loss data of the cross-domain spatiotemporal data is calculated, and the cross-domain spatiotemporal data is adjusted based on the regularized loss data to obtain multi-scene time series data.

[0183] Each module in the aforementioned multi-scenario time-series data generation device under zero-sample conditions can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0184] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational 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 input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for generating multi-scene time-series data under zero-sample conditions. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0185] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0186] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a method for generating multi-scenario time-series data under zero-sample conditions.

[0187] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for generating multi-scenario time-series data under zero-sample conditions.

[0188] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of a method for generating multi-scenario time-series data under zero-sample conditions.

[0189] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0190] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0191] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0192] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating multi-scenario time-series data under zero-sample conditions, characterized in that, The method includes: The city knowledge graph, satellite images, and scene description text are acquired, and the city knowledge graph is encoded to obtain the city structure semantic information. Based on the satellite image and the scene description text, urban heterogeneous semantic information is generated through a multimodal coding model, and the urban structural semantic information and the urban heterogeneous semantic information are aligned to obtain target structural semantic information and target heterogeneous semantic information. Based on the target structural semantic information and the target heterogeneous semantic information, cross-domain spatiotemporal data is generated through a conditional diffusion generation model. The cross-domain spatiotemporal data is then adjusted and processed through a consistency regularization strategy to obtain multi-scenario time-series data.

2. The method according to claim 1, characterized in that, The encoding process of the city knowledge graph to obtain city structural semantic information includes: The city knowledge graph is modeled to obtain a set of city triples, and the city triples are initialized to obtain a city embedding matrix. Based on the city embedding matrix, weighted encoding is performed using a graph attention network to obtain the city structure semantic information.

3. The method according to claim 2, characterized in that, The process of generating heterogeneous semantic information about the city based on the satellite imagery and the scene description text, using a multimodal coding model, includes: The satellite image and the scene description text are mapped into the same embedding space to obtain a standard satellite image and a standard scene description text; The satellite image is divided into individual tile images, and each tile image is encoded into a visual feature vector using a visual encoder. The visual feature vectors are semantically embedded to obtain semantic embedding information corresponding to each tile image, and the sub-descriptive texts related to each domain in the scene description text are encoded to obtain domain embedding information corresponding to each domain. The semantic embedding information corresponding to each tile image and the domain embedding information corresponding to each domain are used as urban heterogeneous semantic information.

4. The method according to claim 1, characterized in that, The alignment process of the urban structural semantic information and the urban heterogeneous semantic information to obtain target structural semantic information and target heterogeneous semantic information includes: The urban structural semantic information and the urban heterogeneous semantic information are divided into regions to obtain the sub-urban structural semantic information and the sub-urban heterogeneous semantic information corresponding to each region. For each region, a semantic reference space for the heterogeneous semantic information of the sub-cities corresponding to the region is identified, and the structural semantic information of the sub-cities is aligned to the semantic reference space through a spatial alignment model to obtain the structural semantic information of the sub-targets. The sub-target structural semantic information of all regions is used as the target structural semantic information, and the heterogeneous semantic information of cities is directly used as the target heterogeneous semantic information.

5. The method according to claim 4, characterized in that, The generation of cross-domain spatiotemporal data based on the target structural semantic information and the target heterogeneous semantic information, through a conditional diffusion generation model, includes: By using a unified temporal coding model, temporal coding is performed on the target structural semantic information and the heterogeneous semantic information to obtain latent structural representation information and latent heterogeneous representation information. The latent structural representation information and the latent heterogeneous representation information are mapped back to the original time domain using a time decoder to obtain the structural reconstruction sequence and the heterogeneous reconstruction sequence. Based on the structural reconstruction sequence and the heterogeneous reconstruction sequence, cross-domain spatiotemporal data are obtained by guiding trajectory generation through a conditional diffusion model.

6. The method according to claim 1, characterized in that, The cross-domain spatiotemporal data is adjusted and processed using a consistency regularization strategy to obtain multi-scenario time-series data, including: Based on the aforementioned cross-domain spatiotemporal data, perturbation procedures are used to construct perturbed conditional embedded spatiotemporal data. Based on the cross-domain spatiotemporal data and the conditionally embedded spatiotemporal data, a noise evaluation procedure is used to generate a first noise output result corresponding to the cross-domain spatiotemporal data and a second noise output result corresponding to the conditionally embedded spatiotemporal data. Based on the first noise output result and the second noise output result, the regularized loss data of the cross-domain spatiotemporal data is calculated, and the cross-domain spatiotemporal data is adjusted based on the regularized loss data to obtain multi-scene time series data.

7. A device for generating multi-scene time-series data under zero-sample conditions, characterized in that, The device includes: The acquisition module is used to acquire urban knowledge graphs, satellite images, and scene description text, and to encode the urban knowledge graphs to obtain urban structural semantic information. The alignment module is used to generate urban heterogeneous semantic information based on the satellite image and the scene description text through a multimodal coding model, and to align the urban structural semantic information and the urban heterogeneous semantic information to obtain target structural semantic information and target heterogeneous semantic information. The adjustment module is used to generate cross-domain spatiotemporal data based on the target structural semantic information and the target heterogeneous semantic information through a conditional diffusion generation model, and to adjust the cross-domain spatiotemporal data through a consistency regularization strategy to obtain multi-scenario time series data.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images