Prediction task processing method and device based on basic model

Abstract

The invention provides a prediction task processing method and device based on a basic model, and the method comprises the steps: obtaining target mobile network data and environment data corresponding to a target prediction task based on a data source corresponding to the target prediction task, and the data source comprises a base station facility layer, a user layer and a wireless channel layer; based on the cue word network of the basic model, performing feature extraction on the target mobile network data and the environment data to obtain a cue vector; based on a time mask mechanism of the basic model and a task type corresponding to the target prediction task, performing mask processing on the target mobile network data to obtain mask mobile network data; and based on a prediction task processing layer of the basic model, performing task processing on the mask mobile network data and the prompt vector to obtain target prediction data corresponding to the target prediction task. According to the invention, different prediction tasks can be executed on different mobile network data, and the accuracy of the prediction result can be improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a prediction task processing method and device based on a basic model. Background Art

[0002] With the rapid development of mobile communication technology, future mobile wireless networks will provide a richer range of services and resources in diverse scenarios, including commuting, production, daily life, and entertainment. Accurately predicting mobile network data (such as base station traffic, user behavior, and channel quality) is crucial for optimizing network efficiency and service quality.

[0003] However, existing mobile data prediction paradigms mostly employ model structures tailored to specific data sources. Different data types exhibit significant differences in temporal granularity, generation mechanisms, and environmental dependencies. Designing a specific model based on a single data source increases overall system complexity and limits generalization capabilities. Therefore, an effective solution is urgently needed to address these issues. Summary of the Invention

[0004] To solve the above problems, the present invention provides a prediction task processing method and device based on a basic model.

[0005] The present invention provides a prediction task processing method based on a basic model, comprising the following steps: Based on a data source corresponding to the target prediction task, target mobile network data and environment data corresponding to the target prediction task are obtained, wherein the data source includes a base station facility layer, a user layer, and a wireless channel layer; Based on the prompt word network of the basic model, feature extraction is performed on the target mobile network data and the environmental data to obtain a prompt vector; Based on the time mask mechanism of the basic model and the task type corresponding to the target prediction task, masking the target mobile network data to obtain masked mobile network data; Based on the prediction task processing layer of the basic model, task processing is performed on the mask movement network data and the prompt vector to obtain target prediction data corresponding to the target prediction task.

[0006] According to a prediction task processing method based on a basic model provided by the present invention, the prompt word network includes a variational autoencoder, a soft prompt learning mechanism and a splicing layer; The prompt word network based on the basic model performs feature extraction processing on the target mobile network data and the environmental data to obtain a prompt vector, including: Based on the variational autoencoder, semantic features are extracted from the environmental data to obtain a semantic space vector; Based on the soft prompt learning mechanism, extracting soft prompt features from the target mobile network data to obtain a soft prompt vector; Based on the concatenation layer, the semantic space vector and the soft hint vector are concatenated to obtain a hint vector.

[0007] According to a prediction task processing method based on a basic model provided by the present invention, the soft prompt learning mechanism includes a periodic prompt mechanism, a time correlation prompt mechanism, and a high-dimensional feature dependency prompt mechanism, and the soft prompt vector includes a main frequency component vector, a temporal correlation feature vector, and a spatial vector; The step of extracting soft prompt features from the target mobile network data based on the soft prompt learning mechanism to obtain a soft prompt vector includes: Based on the periodic prompt mechanism, a fast Fourier transform is used to extract the main frequency component of the target mobile network data to obtain the main frequency component vector; Based on the time correlation prompt mechanism, a Transformer network is used to extract features of the target mobile network data in the time dimension to obtain the temporal correlation feature vector; Based on a high-dimensional feature dependency prompt mechanism, a Transformer network is used to extract features of the target mobile network data in the embedding space dimension to obtain the spatial vector.

[0008] According to a prediction task processing method based on a basic model provided by the present invention, the target mobile network data is masked based on a time mask mechanism of the basic model and a task type corresponding to the target prediction task to obtain masked mobile network data, including: In a case where the task type is a short-term prediction task, masking the target mobile network data based on the short-term mask corresponding to the short-term prediction task in the time mask mechanism to obtain masked mobile network data; In a case where the task type is a long-term prediction task, masking the target mobile network data based on the long-term mask corresponding to the long-term prediction task in the time mask mechanism to obtain masked mobile network data; When the task type is a data generation task, the target mobile network data is masked based on the generation mask corresponding to the data generation task in the time mask mechanism to obtain masked mobile network data, and each bit of the generation mask is 0.

[0009] According to a prediction task processing method based on a basic model provided by the present invention, the prediction task processing layer includes a diffusion model and a Transformer network; The prediction task processing layer based on the basic model performs task processing on the mask movement network data and the hint vector to obtain target prediction data corresponding to the target prediction task, including: Based on the diffusion model, the masked mobile network data is gradually disturbed by adding noise to obtain noisy mobile network data; Based on the Transformer network, the noisy mobile network data is denoised and reconstructed to obtain target prediction data corresponding to the target prediction task.

[0010] According to a basic model-based prediction task processing method provided by the present invention, the basic model-based prompt word network extracts features from the target mobile network data and the environmental data to obtain a prompt vector, further comprising: The target mobile network data is segmented using a temporal convolutional network to obtain segmented target mobile network data.

[0011] According to a basic model-based prediction task processing method provided by the present invention, the basic model-based prompt word network extracts features from the target mobile network data and the environmental data to obtain a prompt vector, further comprising: Constructing a training set based on different task types, sample mobile network data from different data sources and their corresponding sample environment data, wherein the training set includes multiple sample task sets, each of which includes sample tasks, sample mobile network data and their corresponding sample environment data; Model training is performed on an untrained basic model based on each of the sample task sets to obtain a trained basic model; the loss function of the model training is a self-supervised loss function.

[0012] The present invention also provides a prediction task processing device based on the basic model, comprising the following modules: an acquisition module configured to acquire target mobile network data and environment data corresponding to the target prediction task based on a data source corresponding to the target prediction task, the data source including a base station facility layer, a user layer, and a wireless channel layer; a feature extraction module configured to extract features from the target mobile network data and the environment data based on a prompt word network of a basic model to obtain a prompt vector; a masking module configured to perform masking on the target mobile network data based on the time masking mechanism of the basic model and the task type corresponding to the target prediction task to obtain masked mobile network data; The task processing module is configured to perform task processing on the mask movement network data and the prompt vector based on the prediction task processing layer of the basic model to obtain target prediction data corresponding to the target prediction task.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the prediction task processing method based on the basic model as described above is implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described prediction task processing methods based on the basic model.

[0015] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned basic model-based prediction task processing methods.

[0016] The present invention provides a prediction task processing method and device based on a basic model. The method and device obtain target mobile network data and environmental data corresponding to the target prediction task based on the data source corresponding to the target prediction task, and the data source includes a base station facility layer, a user layer, and a wireless channel layer; based on the prompt word network of the basic model, feature extraction is performed on the target mobile network data and the environmental data to obtain a prompt vector; based on the time mask mechanism of the basic model and the task type corresponding to the target prediction task, mask processing is performed on the target mobile network data to obtain masked mobile network data; based on the prediction task processing layer of the basic model, task processing is performed on the masked mobile network data and the prompt vector to obtain target prediction data corresponding to the target prediction task. The present invention can not only perform different prediction tasks on different mobile network data, but also improve the accuracy of the prediction results, solving the problems of model fragmentation, task isolation, and insufficient environmental adaptability in existing mobile network data prediction schemes. Among them, the basic model, while maintaining a high degree of versatility, has the ability to adaptively generalize to multiple tasks, multiple types, and multiple scenarios. It can support short-term predictions, long-term trend predictions, and distributed generation tasks in scenarios without historical data. It is suitable for traffic analysis, resource scheduling, and digital twin simulation in real network environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 It is a flowchart of the prediction task processing method based on the basic model provided by the present invention.

[0019] Figure 2 It is a schematic diagram of the application scenario provided by the present invention.

[0020] Figure 3 It is a structural diagram of the basic model provided by the present invention.

[0021] Figure 4 It is a schematic diagram of the task processing result provided by the present invention.

[0022] Figure 5 It is a structural diagram of the prediction task processing device based on the basic model provided by the present invention.

[0023] Figure 6 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0025] The following combination Figures 1-6 The present invention describes a method and device for processing prediction tasks based on a basic model.

[0026] First, the relevant contents of the present invention are briefly described.

[0027] In mobile networks, data is diverse, including base station traffic data at the infrastructure layer, application (Application) usage behavior at the user layer, and channel quality parameters at the wireless channel layer, such as Reference Signal Received Power (RSRP).

[0028] Through forward-looking predictions of the above data, operators can proactively perceive changes in network status, timely dispatch wireless resources, and rationally arrange network infrastructure and user resource allocation, thereby improving information supply efficiency and user service experience.

[0029] Existing mobile data prediction paradigms mostly employ model structures tailored to specific data sources, specifically designing specialized models for base station traffic, user application usage, or channel quality. For example, Transformers are combined with Graph Convolutional Networks (GCNs) to extract the spatiotemporal dependencies of mobile traffic sequences for efficient network traffic generation. For example, a frequency attention mechanism is employed to generate traffic using a diffusion model. For example, a model based on PacketCGAN (a data balancing and classification method based on generative adversarial networks) is used to address data imbalance in databases covering multiple smartphone service types. For example, Long Short-Term Memory (LSTM) networks are used to directly learn temporal features from channel data to achieve channel quality estimation. For example, features are manually constructed from base station transmission engineering data to predict RSRP, including path loss calculation and multipath signal angle calculation.

[0030] Although this type of "customized" strategy can achieve good results in a single scenario, in a large-scale network environment that includes multi-dimensional heterogeneous elements such as base stations, users and channels, it often significantly increases the complexity of model design and deployment costs, which is not conducive to the promotion and application of universal models.

[0031] In recent years, large-scale foundation models have rapidly emerged in the fields of computer vision and natural language processing. With their powerful data processing capabilities, excellent generalization performance, and zero- and small-shot learning capabilities, they are reshaping the development landscape of artificial intelligence. Multiple vertical industries (such as urban planning, finance, and healthcare) have begun exploring solutions for building foundation models based on domain knowledge. Some research even proposes the concept of "world models" to achieve artificial general intelligence (AGI). Research demonstrates that foundation models not only possess the ability to perceive real-world data but also predict future data, providing solid support for data-driven decision-making.

[0032] As the infrastructure that carries massive amounts of information resources and service capabilities, mobile wireless networks exhibit typical multidimensional, heterogeneous, and time-sensitive characteristics. Future mobile networks will further develop towards integrated integration across land, air, and sea. The number of devices and service complexity will continue to grow, and the scale of network data will increase exponentially. Basic models have broad application potential in this context. On the one hand, basic models can accurately capture the dynamic evolution of mobile data, assisting operators in proactive optimization and intelligent decision-making. On the other hand, their zero-sample and small-sample capabilities enable service demand estimation in areas lacking historical data, helping to improve network planning efficiency. Furthermore, basic models can extract behavioral distribution patterns without relying on private user data, potentially balancing data modeling accuracy with privacy protection requirements. Furthermore, basic models can be used to construct digital twins of mobile networks, synthesizing multi-layered network data in a virtual environment and enabling comparative simulations under different engineering parameters. This provides strong support for operational strategy evaluation (e.g., cell physical signage planning and user mobility management). Currently, there are no general-purpose basic models specifically designed for mobile wireless networks, and related research remains in the exploratory stage, leaving significant technical gaps and room for development.

[0033] Figure 1 This is a flow chart of the prediction task processing method based on the basic model provided by the present invention, such as Figure 1 As shown, the method includes steps 101 to 104.

[0034] Step 101: Based on a data source corresponding to a target prediction task, target mobile network data and environmental data corresponding to the target prediction task are obtained, where the data source includes a base station facility layer, a user layer, and a wireless channel layer.

[0035] Specifically, a target prediction task refers to a task that predicts mobile network data. The task type of a target prediction task can be any of the following: a short-term prediction task, a long-term prediction task, or a data generation task. Furthermore, the task type can also be a random task, which can be any task other than a short-term prediction task, a long-term prediction task, or a data generation task, such as a combination of at least two of these tasks.

[0036] See also Figure 2 , Figure 2 This is a schematic diagram of the application scenario provided by the present invention. The prediction task processing method based on the basic model provided by the present invention can be applied to mobile network data from different data sources. The data sources include the base station facility layer (mobile infrastructure), the user layer (users), and the wireless channel layer (wireless channels). The mobile network data corresponding to the base station facility layer is the base station traffic sequence. (Base station / cell traffic); the mobile network data corresponding to the user layer is the App usage traffic sequence (User Apps use); The mobile network data corresponding to the wireless channel layer is the downlink RSRP sequence (Wireless channel). Mobile network data from different data sources all have a time series structure. Base station traffic data is sampled at hourly granularity, app usage data at second-level granularity, and downlink RSRP data at millisecond-level granularity. These three data types exhibit significant temporal heterogeneity.

[0037] Different data sources correspond to different environmental data, which can be environmental condition characteristics. Specifically, the environmental data corresponding to the base station facility layer is urban space information. , such as information point (POI) density and base station topology; the environmental data corresponding to the user layer is the user portrait feature , such as application preferences and usage habits; the environmental data corresponding to the wireless channel layer is the engineering parameters , such as antenna tilt angle, transmission power, propagation path, geographical type, etc.

[0038] In practical applications, a target prediction task may be received first, then analyzed to determine a data source corresponding to the target prediction task, and then corresponding target mobile network data may be obtained from the data source.

[0039] For example, if the target prediction task is short-term network load prediction, the corresponding data sources are the base station facility layer and the user layer, and the base station traffic sequence and the App usage traffic sequence can be obtained as the target mobile network data.

[0040] Step 102: Based on the prompt word network of the basic model, feature extraction is performed on the target mobile network data and the environment data to obtain a prompt vector.

[0041] In practical applications, after obtaining the target mobile network data and environmental data corresponding to the target prediction task, the target mobile network data and environmental data corresponding to the target prediction task can be input into the basic model.

[0042] See also Figure 3 , Figure 3It is a structural diagram of the basic model provided by the present invention. The input data of the basic model can be multi-type massive mobile data (mobile network data corresponding to different data sources) and domain environment data (environmental data corresponding to different data sources). Among them, multi-type massive mobile data includes base station traffic sequence (base station / cell traffic); the mobile network data corresponding to the user layer is the App usage traffic sequence (user APP usage); the mobile network data corresponding to the wireless channel layer is the downlink RSRP sequence (wireless channel, that is, channel quality (RSRP)). Domain environment data includes urban space information (urban environment), user portrait features (user portrait) and engineering parameters.

[0043] Further, see Figure 3 ,The multi-domain prompt word network in the basic model processes the target ,mobile network data and environmental data to obtain the prompt vector of the basic ,model.

[0044] Step 103: Based on the time mask mechanism of the basic model and the task type corresponding to the target prediction task, mask processing is performed on the target mobile network data to obtain masked mobile network data.

[0045] Specifically, the masked mobile network data refers to the mobile network data after data masking.

[0046] In practical applications, after inputting the target mobile network data corresponding to the target prediction task into the basic model, see Figure 3 ,The multi-task oriented temporal mask mechanism in the basic model performs mask processing on the target ,mobile network data, i.e. mobile data embedding, thereby ,masking the mobile network data.

[0047] Step 104: Based on the prediction task processing layer of the basic model, task processing is performed on the mask movement network data and the prompt vector to obtain target prediction data corresponding to the target prediction task.

[0048] In actual applications, after obtaining the mask movement network data and the prompt vector, the prediction task processing layer of the basic model performs prediction based on the mask movement network data and the prompt vector, thereby obtaining the target prediction data corresponding to the target prediction task.

[0049] The present invention provides a prediction task processing method based on a basic model. The method obtains target mobile network data and environmental data corresponding to the target prediction task based on a data source corresponding to the target prediction task, the data source including a base station facility layer, a user layer, and a wireless channel layer; performs feature extraction on the target mobile network data and the environmental data based on a prompt word network of the basic model to obtain a prompt vector; performs mask processing on the target mobile network data based on a time mask mechanism of the basic model and a task type corresponding to the target prediction task to obtain masked mobile network data; performs task processing on the masked mobile network data and the prompt vector based on a prediction task processing layer of the basic model to obtain target prediction data corresponding to the target prediction task. The present invention can not only perform different prediction tasks on different mobile network data, but also improve the accuracy of the prediction results, solving the problems of model fragmentation, task isolation, and insufficient environmental adaptability in existing mobile network data prediction schemes. Among them, the basic model, while maintaining a high degree of versatility, has the ability to adaptively generalize to multiple tasks, multiple types, and multiple scenarios. It can support short-term predictions, long-term trend predictions, and distributed generation tasks in scenarios without historical data. It is suitable for traffic analysis, resource scheduling, and digital twin simulation in real network environments.

[0050] Optionally, the prompt word network includes a variational autoencoder, a soft prompt learning mechanism, and a splicing layer; accordingly, the prompt word network based on the basic model performs feature extraction processing on the target mobile network data and the environment data to obtain a prompt vector, including: Based on the variational autoencoder, semantic features are extracted from the environmental data to obtain a semantic space vector; Based on the soft prompt learning mechanism, extracting soft prompt features from the target mobile network data to obtain a soft prompt vector; Based on the concatenation layer, the semantic space vector and the soft hint vector are concatenated to obtain a hint vector.

[0051] Specifically, the soft prompt learning mechanism refers to the task adaptation mechanism of soft prompt learning (Soft Prompt Learning with Task Awareness).

[0052] In practical applications, in order to enable the basic model to have the ability to recognize different types of mobile data structures and semantics, a variational autoencoder, a soft prompt learning mechanism and a splicing layer are set up in the prompt word network.

[0053] See also Figure 3In a multi-domain prompt word network, a variational autoencoder (VAE) can be used to uniformly compress and map the context data into a latent space, i.e., semantic embedding, to obtain the VAE output corresponding to the context data, i.e., the context embedding representation. The VAE processing process is as follows: in, is the environment embedding representation, Represents any environmental data among urban space information, user portrait characteristics and engineering parameters. represents the mean, represents the variance, represents random noise, represents the standard normal distribution.

[0054] Then, the environment embedding representation is unified into a semantic space vector , that is, semantic extraction. represents the batch size, is the length of time, This allows for alignment of mobile network data from different data sources at the context semantic layer.

[0055] Also, see Figure 3 ,In the multi-domain oriented prompt word network, soft prompt features can be ,extracted from the target mobile network data based on the soft prompt ,learning mechanism, thereby obtaining learnable parameters, i.e. soft prompt vectors.

[0056] Furthermore, in a multi-domain prompt word network, the semantic space vector and the soft prompt vector can be concatenated based on the concatenation layer to obtain the prompt vector of the basic model. This can improve the accuracy of the prompt vector.

[0057] In one or more optional embodiments of the present invention, the soft prompt learning mechanism includes a periodic prompt mechanism, a time correlation prompt mechanism, and a high-dimensional feature dependency prompt mechanism, and the soft prompt vector includes a main frequency component vector, a temporal correlation feature vector, and a spatial vector; The step of extracting soft prompt features from the target mobile network data based on the soft prompt learning mechanism to obtain a soft prompt vector includes: Based on the periodic prompt mechanism, a fast Fourier transform is used to extract the main frequency component of the target mobile network data to obtain the main frequency component vector; Based on the time correlation prompt mechanism, a Transformer network is used to extract features of the target mobile network data in the time dimension to obtain the temporal correlation feature vector; Based on a high-dimensional feature dependency prompt mechanism, a Transformer network is used to extract features of the target mobile network data in the embedding space dimension to obtain the spatial vector.

[0058] In practical applications, three categories of soft hint vectors are injected into the model input, especially the target mobile network data, to learn periodicity, time correlation and high-dimensional feature dependency information in a data-driven manner.

[0059] First, in terms of periodic prompts, the main frequency component vector of the target mobile network data is extracted using fast Fourier transform : in, represents the Fourier transform operator, Indicates the selection of the strongest frequency components, Indicates the first k bits of data.

[0060] Secondly, in terms of temporal correlation prompts, the Transformer network is used to extract features of the sequence in the time dimension to obtain the temporal correlation feature vector : in, Represents a Transformer module that models along the time dimension.

[0061] Then, in terms of high-dimensional feature dependency hints, the Transformer network is used to model the relationship between features in the embedding space dimension, that is, to perform feature extraction and obtain the spatial vector : in, Represents the relationships between features modeled in the embedding space dimensions.

[0062] are presented in the form of learnable parameters. Finally, the prompt vector of the base model is the semantic space vector Concatenation with three learnable parameters: Through the prompt mechanism, the model is guided to perceive different data types and task scenarios, enhancing generalization and migration capabilities.

[0063] In one or more optional embodiments of the present invention, the masking process is performed on the target mobile network data based on the time masking mechanism of the basic model and the task type corresponding to the target prediction task to obtain the masked mobile network data, including: In a case where the task type is a short-term prediction task, masking the target mobile network data based on the short-term mask corresponding to the short-term prediction task in the time mask mechanism to obtain masked mobile network data; In a case where the task type is a long-term prediction task, masking the target mobile network data based on the long-term mask corresponding to the long-term prediction task in the time mask mechanism to obtain masked mobile network data; When the task type is a data generation task, the target mobile network data is masked based on the generation mask corresponding to the data generation task in the time mask mechanism to obtain masked mobile network data, and each bit of the generation mask is 0.

[0064] In practical applications, a differentiated temporal masking mechanism (TMM) was designed to uniformly model the three core tasks of short-term prediction, long-term prediction, and data generation. This multi-task TMM guides the model to perform task differentiation and target alignment on input data through supervised and self-supervised approaches.

[0065] In short-term forecasting tasks, the base model uses known historical data fragments (target mobile network data) to predict near-future data (target prediction data), corresponding short-term mask for: in, l The number of items or bits of data contained in the target mobile network data. The total number of items or bits of data included in the target prediction data and the target mobile network data.

[0066] In long-term prediction tasks, the basic model needs to infer future data (target prediction data) with long-term trends based on shorter historical data (target mobile network data), that is, , i.e. long-term mask With short-term mask The mask structure is similar to that of , but the masked area is longer.

[0067] In the data generation task, the basic model does not rely on any historical input, but is completely based on environmental data for data distribution modeling, and adopts a full masking strategy for the entire input time dimension (target mobile network data), that is, generating a mask for: In addition, in order to enhance the model's ability to perceive temporal correlation, the present invention introduces a random masking mechanism (RandomMasking) to perform random position masking on the time dimension to improve the model's reconstruction and generalization capabilities. That is, when the prediction task is a random task, a random mask is used to mask the target mobile network data to obtain masked mobile network data.

[0068] See also Figure 3 ,In the multi-task oriented temporal mask mechanism, the mask operation on the target ,mobile network data is performed through random mask, short-term mask, long-term mask, or ,generated mask time, namely mobile data embedding.

[0069] In one or more optional embodiments of the present invention, the prediction task processing layer includes a diffusion model and a Transformer network; The prediction task processing layer based on the basic model performs task processing on the mask movement network data and the hint vector to obtain target prediction data corresponding to the target prediction task, including: Based on the diffusion model, the masked mobile network data is gradually disturbed by adding noise to obtain noisy mobile network data; Based on the Transformer network, the noisy mobile network data is denoised and reconstructed to obtain target prediction data corresponding to the target prediction task.

[0070] In practical applications, the diffusion model is combined with the Transformer architecture to obtain a prediction task processing layer for modeling and predicting target mobile network data. By adding noise to the original sequence (masked mobile network data), the original sequence (masked mobile network data) is gradually perturbed, such as by introducing step-by-step Gaussian noise perturbation (Noise Perturbation), to obtain the noisy mobile network data, i.e. Figure 3 In this way, the generation and reconstruction capabilities of the diffusion model can be improved.

[0071] Specifically, the simulated diffusion process is as follows: in, represents the cumulative weight of the diffusion step, Indicates the i The weight of the step diffusion,k Indicates the k Step diffusion, To mask mobile network data, Indicates the k The noisy mobile network data obtained by step diffusion, represents the standard normal distribution, is the standard normal noise.

[0072] See also Figure 3 In the reverse generation process, the base model uses a conditionally conditioned Transformer network (Transformer block) to perform denoising and reconstruction to obtain the target prediction data, that is, denoising and reconstruction is performed through the denoising network. The denoising network includes N Transformer blocks (Transformer block × N) and an output encoder. The guidance and control of the conditional information are achieved by adjusting the layer normalization parameters: in, and They are attention layer and multi-layer perceptron, is the layer normalization function, is the noisy embedding representation, 、 It is the intermediate representation updated layer by layer in the Transformer block. It is a scaling factor that controls the output amplitude of the attention module and the feedforward module in the inter-layer residual path. It is the affine transformation parameter in Condition Scaling, which is obtained by inputting the environment data and prompt vector into the multilayer perceptron.

[0073] In one or more optional implementations of the present invention, the basic model-based prompt word network, before extracting features from the target mobile network data and the environment data to obtain a prompt vector, further includes: The target mobile network data is segmented using a temporal convolutional network to obtain segmented target mobile network data.

[0074] In practical applications, in order to further adapt multiple types of mobile data to a unified time series modeling structure, a tokenization mechanism is introduced when inputting target network data into the basic model.

[0075] As shown in Figure 3, the target network data is segmented through a temporal convolution network to extract its latent space representation (target mobile network data after segmentation). ,in represents the batch size, is the length of time, is the embedding dimension. This process is equivalent to dividing the continuous time series into learnable semantic units, making it more suitable for the Transformer structure.

[0076] In one or more optional embodiments of the present invention, before the basic model-based prompt word network extracts features from the target mobile network data and the environment data to obtain the prompt vector, the method further includes: Constructing a training set based on different task types, sample mobile network data from different data sources and their corresponding sample environment data, wherein the training set includes multiple sample task sets, each of which includes sample tasks, sample mobile network data and their corresponding sample environment data; Model training is performed on an untrained basic model based on each of the sample task sets to obtain a trained basic model; the loss function of the model training is a self-supervised loss function.

[0077] In actual applications, sample task sets corresponding to different task types and different data sources can be obtained, such as short-term prediction tasks, long-term prediction tasks, and data generation tasks for base station traffic sequences, short-term prediction tasks, long-term prediction tasks, and data generation tasks for App usage traffic sequences, and short-term prediction tasks, long-term prediction tasks, and data generation tasks for downlink RSRP sequences.

[0078] Then, the untrained basic model is trained self-supervised based on these sample task sets, where the self-supervised loss function is as follows: in, is the actual injected Gaussian noise (label). The noise predicted by the model is used as the output target. is the conditional cue vector generated by the task type (prediction / generation) and the semantic cue network. is the task mask. This loss guides the model to recover the true sequence from the noisy input, improving its performance in prediction and generation tasks.

[0079] The following describes the prediction task processing method based on the basic model provided by the present invention in conjunction with specific scenarios.

[0080] In actual use, the input of the basic model includes three types of mobile network data: base station traffic data at the infrastructure layer (base station traffic sequence), application usage data at the user layer (app usage traffic sequence), and RSRP data at the wireless channel layer (downlink RSRP sequence), such as Figure 2These data first go through a unified preprocessing process, which mainly includes time alignment, normalization, and slice window construction, to form fixed-length time series segments for the basic model to receive.

[0081] Among them, time alignment refers to synchronizing data with different sampling frequencies to a unified time base to ensure that the model receives valid corresponding values ​​of various types of data at the same time. Normalization is used to scale different data dimensions (such as traffic size, signal strength, etc.) to a similar numerical range to avoid interference from numerical differences in model training. Slice window construction is to divide the continuous time series into sliding window segments of fixed length. Each segment serves as a complete input sample for the model to learn time series patterns and changing trends. Environmental information is provided by the operator's existing infrastructure data and Geographic Information System (GIS) data, including base station deployment parameters, urban area functions, user attributes, antenna configuration, etc. Environmental data is encoded as a vector and input into the model to support conditional modeling.

[0082] During model inference, users can specify the type of prediction task. For example, to perform short-term network load forecasting, users can input a sequence of base station traffic and user data from the last few hours, and the model will output traffic forecasts for the next period. To perform long-term network evolution assessments, users only need to input a short period of time (traffic sequence) in the past, and the model will predict trend data for a longer period in the future. To simulate network usage in a new area, users only need to provide the area's environmental parameters in the absence of historical data, and the model will generate simulated network data for that area, which can be used for digital twins or site deployment assessments.

[0083] Figure 4 This is a schematic diagram of the task processing results provided by the present invention, see Figure 4 , Figure 4 This visualization shows the results of short-term prediction (short-term prediction task), long-term prediction (long-term prediction task), and distribution generation (data generation task). The basic model achieves accurate prediction and generation. The horizontal axis represents time (hours), and the vertical axis represents mobile traffic data (mobile traffic date). The solid blue line represents the predicted data (generated date), and the dashed red line represents the real data (real date).

[0084] Furthermore, the basic model supports both local and cloud deployment. Operators can perform real-time predictions on edge servers or encapsulate the model as an inference service application programming interface (API) to access traffic, usage behavior, or RSRP predictions via the Hypertext Transfer Protocol (HTTP) interface. The entire system is suitable for scenarios including urban network planning, network load balancing, site deployment optimization, and virtual network experiments, offering strong feasibility, high task versatility, and flexible deployment.

[0085] This embodiment of the present invention integrates a diffusion generation model with a Transformer network architecture, achieving unified predictive modeling capabilities for diverse mobile data through soft-hint learning, a temporal masking mechanism, and a conditional scaling network. This allows the underlying model to maintain high versatility while also possessing adaptive generalization capabilities for multiple tasks, types, and scenarios. It supports short-term and long-term trend forecasting, as well as distributed generation tasks in scenarios without historical data. It is suitable for traffic analysis, resource scheduling, and digital twin simulation in real-world network environments.

[0086] The following describes a prediction task processing device based on a basic model provided by the present invention. The prediction task processing device based on a basic model described below and the prediction task processing method based on a basic model described above can refer to each other.

[0087] Figure 5 This is a schematic diagram of the structure of the prediction task processing device based on the basic model provided by the present invention. Figure 5 As shown, the device includes: An acquisition module 501 is configured to acquire target mobile network data and environment data corresponding to the target prediction task based on a data source corresponding to the target prediction task, where the data source includes a base station facility layer, a user layer, and a wireless channel layer; A feature extraction module 502 is configured to extract features from the target mobile network data and the environment data based on a prompt word network of a basic model to obtain a prompt vector; The masking module 503 is configured to perform masking processing on the target mobile network data based on the time masking mechanism of the basic model and the task type corresponding to the target prediction task to obtain masked mobile network data; The task processing module 504 is configured to perform task processing on the mask movement network data and the hint vector based on the prediction task processing layer of the basic model to obtain target prediction data corresponding to the target prediction task.

[0088] The basic model-based prediction task processing device provided by this invention not only performs different prediction tasks for different mobile network data but also improves the accuracy of prediction results, resolving issues such as model fragmentation, task isolation, and insufficient environmental adaptability that exist in existing mobile network data prediction solutions. While maintaining high versatility, the basic model possesses adaptive generalization capabilities for multiple tasks, types, and scenarios. It supports short-term predictions, long-term trend predictions, and distribution generation tasks in scenarios without historical data. It is suitable for traffic analysis, resource scheduling, and digital twin simulation in real-world network environments.

[0089] In one or more optional embodiments of the present invention, the prompt word network includes a variational autoencoder, a soft prompt learning mechanism, and a splicing layer; The feature extraction module 502 is specifically configured to: Based on the variational autoencoder, semantic features are extracted from the environmental data to obtain a semantic space vector; Based on the soft prompt learning mechanism, extracting soft prompt features from the target mobile network data to obtain a soft prompt vector; Based on the concatenation layer, the semantic space vector and the soft hint vector are concatenated to obtain a hint vector.

[0090] In one or more optional embodiments of the present invention, the soft prompt learning mechanism includes a periodic prompt mechanism, a time correlation prompt mechanism, and a high-dimensional feature dependency prompt mechanism, and the soft prompt vector includes a main frequency component vector, a temporal correlation feature vector, and a spatial vector; The feature extraction module 502 is specifically configured to: Based on the periodic prompt mechanism, a fast Fourier transform is used to extract the main frequency component of the target mobile network data to obtain the main frequency component vector; Based on the time correlation prompt mechanism, a Transformer network is used to extract features of the target mobile network data in the time dimension to obtain the temporal correlation feature vector; Based on a high-dimensional feature dependency prompt mechanism, a Transformer network is used to extract features of the target mobile network data in the embedding space dimension to obtain the spatial vector.

[0091] In one or more optional embodiments of the present invention, the mask module 503 is specifically configured to: In a case where the task type is a short-term prediction task, masking the target mobile network data based on the short-term mask corresponding to the short-term prediction task in the time mask mechanism to obtain masked mobile network data; In a case where the task type is a long-term prediction task, masking the target mobile network data based on the long-term mask corresponding to the long-term prediction task in the time mask mechanism to obtain masked mobile network data; When the task type is a data generation task, the target mobile network data is masked based on the generation mask corresponding to the data generation task in the time mask mechanism to obtain masked mobile network data, and each bit of the generation mask is 0.

[0092] In one or more optional embodiments of the present invention, the prediction task processing layer includes a diffusion model and a Transformer network; The task processing module 504 is specifically configured to: Based on the diffusion model, the masked mobile network data is gradually disturbed by adding noise to obtain noisy mobile network data; Based on the Transformer network, the noisy mobile network data is denoised and reconstructed to obtain target prediction data corresponding to the target prediction task.

[0093] In one or more optional implementations of the present invention, the apparatus further includes a word segmentation module configured to: The target mobile network data is segmented using a temporal convolutional network to obtain segmented target mobile network data.

[0094] In one or more optional implementations of the present invention, the apparatus further includes a training module configured to: Constructing a training set based on different task types, sample mobile network data from different data sources and their corresponding sample environment data, wherein the training set includes multiple sample task sets, each of which includes sample tasks, sample mobile network data and their corresponding sample environment data; Model training is performed on an untrained basic model based on each of the sample task sets to obtain a trained basic model; the loss function of the model training is a self-supervised loss function.

[0095] Figure 6 An example of a physical structure diagram of an electronic device is shown below. Figure 6As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communications bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other via the communications bus 640. The processor 610 may call logic instructions in the memory 630 to execute a prediction task processing method based on a basic model. The method includes: obtaining target mobile network data and environmental data corresponding to the target prediction task based on a data source corresponding to the target prediction task, the data source including a base station facility layer, a user layer, and a wireless channel layer; performing feature extraction on the target mobile network data and the environmental data based on a prompt word network of the basic model to obtain a prompt vector; performing masking processing on the target mobile network data based on a time masking mechanism of the basic model and a task type corresponding to the target prediction task to obtain masked mobile network data; and performing task processing on the masked mobile network data and the prompt vector based on a prediction task processing layer of the basic model to obtain target prediction data corresponding to the target prediction task.

[0096] Furthermore, the logic instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0097] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the prediction task processing method based on the basic model provided by the above methods, the method including: based on the data source corresponding to the target prediction task, obtaining the target mobile network data and environmental data corresponding to the target prediction task, the data source including the base station facility layer, the user layer and the wireless channel layer; based on the prompt word network of the basic model, performing feature extraction on the target mobile network data and the environmental data to obtain a prompt vector; based on the time mask mechanism of the basic model and the task type corresponding to the target prediction task, performing mask processing on the target mobile network data to obtain masked mobile network data; based on the prediction task processing layer of the basic model, performing task processing on the masked mobile network data and the prompt vector to obtain the target prediction data corresponding to the target prediction task.

[0098] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the prediction task processing method based on the basic model provided by the above-mentioned methods, the method comprising: based on the data source corresponding to the target prediction task, obtaining the target mobile network data and environmental data corresponding to the target prediction task, the data source including the base station facility layer, the user layer and the wireless channel layer; based on the prompt word network of the basic model, performing feature extraction on the target mobile network data and the environmental data to obtain a prompt vector; based on the time mask mechanism of the basic model and the task type corresponding to the target prediction task, performing mask processing on the target mobile network data to obtain masked mobile network data; based on the prediction task processing layer of the basic model, performing task processing on the masked mobile network data and the prompt vector to obtain the target prediction data corresponding to the target prediction task.

[0099] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0100] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A prediction task processing method based on a basic model, characterized in that: include: Based on a data source corresponding to the target prediction task, target mobile network data and environment data corresponding to the target prediction task are obtained, wherein the data source includes a base station facility layer, a user layer, and a wireless channel layer; Based on the prompt word network of the basic model, feature extraction is performed on the target mobile network data and the environmental data to obtain a prompt vector; Based on the time mask mechanism of the basic model and the task type corresponding to the target prediction task, masking the target mobile network data to obtain masked mobile network data; Based on the prediction task processing layer of the basic model, task processing is performed on the mask movement network data and the prompt vector to obtain target prediction data corresponding to the target prediction task.

2. The prediction task processing method based on the basic model according to claim 1 is characterized in that: The prompt word network includes a variational autoencoder, a soft prompt learning mechanism and a splicing layer; The prompt word network based on the basic model performs feature extraction processing on the target mobile network data and the environmental data to obtain a prompt vector, including: Based on the variational autoencoder, semantic features are extracted from the environmental data to obtain a semantic space vector; Based on the soft prompt learning mechanism, extracting soft prompt features from the target mobile network data to obtain a soft prompt vector; Based on the concatenation layer, the semantic space vector and the soft hint vector are concatenated to obtain a hint vector.

3. The prediction task processing method based on the basic model according to claim 2, characterized in that: The soft prompt learning mechanism includes a periodic prompt mechanism, a time correlation prompt mechanism and a high-dimensional feature dependency prompt mechanism, and the soft prompt vector includes a main frequency component vector, a temporal correlation feature vector and a spatial vector; The step of extracting soft prompt features from the target mobile network data based on the soft prompt learning mechanism to obtain a soft prompt vector includes: Based on the periodic prompt mechanism, a fast Fourier transform is used to extract the main frequency component of the target mobile network data to obtain the main frequency component vector; Based on the time correlation prompt mechanism, a Transformer network is used to extract features of the target mobile network data in the time dimension to obtain the temporal correlation feature vector; Based on a high-dimensional feature dependency prompt mechanism, a Transformer network is used to extract features of the target mobile network data in the embedding space dimension to obtain the spatial vector.

4. The prediction task processing method based on the basic model according to claim 1 is characterized in that: The time masking mechanism based on the basic model and the task type corresponding to the target prediction task, performing masking processing on the target mobile network data to obtain masked mobile network data, includes: In a case where the task type is a short-term prediction task, masking the target mobile network data based on the short-term mask corresponding to the short-term prediction task in the time mask mechanism to obtain masked mobile network data; In a case where the task type is a long-term prediction task, masking the target mobile network data based on the long-term mask corresponding to the long-term prediction task in the time mask mechanism to obtain masked mobile network data; When the task type is a data generation task, the target mobile network data is masked based on the generation mask corresponding to the data generation task in the time mask mechanism to obtain masked mobile network data, and each bit of the generation mask is 0.

5. The prediction task processing method based on the basic model according to claim 1 is characterized in that: The prediction task processing layer includes a diffusion model and a Transformer network; The prediction task processing layer based on the basic model performs task processing on the mask movement network data and the hint vector to obtain target prediction data corresponding to the target prediction task, including: Based on the diffusion model, the masked mobile network data is gradually disturbed by adding noise to obtain noisy mobile network data; Based on the Transformer network, the noisy mobile network data is denoised and reconstructed to obtain target prediction data corresponding to the target prediction task.

6. The prediction task processing method based on the basic model according to claim 1 is characterized in that: Before the prompt word network based on the basic model extracts features from the target mobile network data and the environment data to obtain a prompt vector, the method further includes: The target mobile network data is segmented using a temporal convolutional network to obtain segmented target mobile network data.

7. The method for processing prediction tasks based on a basic model according to any one of claims 1 to 6, characterized in that: Before the prompt word network based on the basic model extracts features from the target mobile network data and the environment data to obtain a prompt vector, the method further includes: Constructing a training set based on different task types, sample mobile network data from different data sources and their corresponding sample environment data, wherein the training set includes multiple sample task sets, each of which includes sample tasks, sample mobile network data and their corresponding sample environment data; Model training is performed on an untrained basic model based on each of the sample task sets to obtain a trained basic model; the loss function of the model training is a self-supervised loss function.

8. A prediction task processing device based on a basic model, characterized in that: include: an acquisition module configured to acquire target mobile network data and environment data corresponding to the target prediction task based on a data source corresponding to the target prediction task, the data source including a base station facility layer, a user layer, and a wireless channel layer; a feature extraction module configured to extract features from the target mobile network data and the environment data based on a prompt word network of a basic model to obtain a prompt vector; a masking module configured to perform masking on the target mobile network data based on the time masking mechanism of the basic model and the task type corresponding to the target prediction task to obtain masked mobile network data; The task processing module is configured to perform task processing on the mask movement network data and the prompt vector based on the prediction task processing layer of the basic model to obtain target prediction data corresponding to the target prediction task.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the prediction task processing method based on the basic model as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for processing prediction tasks based on a basic model according to any one of claims 1 to 7 is implemented.

Images