Time prediction system, method and device based on adaptive prefix generation
By introducing the adaptive prefix generating mechanism and data conversion technology in the time series prediction model, the existing models are not enough in terms of adaptability, versatility and predictive accuracy, and achieved more efficient and flexible time sequence predictions.
Patent Information
- Application Number
- CN202411330035.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-24
- Publication Date
- 2025-01-28
AI Technical Summary
The existing time series prediction models have shortcomings in adaptability, versatility and prediction accuracy, and it is difficult to provide ideal prediction results in diverse application scenarios.
By introducing the adaptive prefix generation mechanism, the time sequence data is converted into the text prototypes that can be processed by a language model, and pre -training language models are used to predict, and the prefix with dynamic generation and specific tasks and context information.
The performance of the time sequence predictive model is significantly improved, the accuracy and robustness of the prediction is improved, the generality and adaptability of the model is enhanced, and the prediction results with more targeted and interpretability are provided.
Smart Images
Figure CN119377288A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of time series prediction, and in particular to a time prediction system, method and device based on adaptive prefix generation. Background Art
[0002] Time series forecasting is widely used in practical applications, such as demand planning, inventory management, energy load forecasting, and climate modeling. Traditional time series forecasting methods mainly include statistical models and machine learning models. Statistical models such as autoregressive moving average models and exponential smoothing methods use historical data of time series for linear prediction, but they have limited performance when dealing with complex nonlinear relationships. Machine learning models such as long short-term memory networks (LSTM) and temporal convolutional networks (TCN) can capture complex patterns and long-term dependencies in time series, but the training process requires a lot of data and computing resources.
[0003] In recent years, pre-trained large language models such as GPT-3, BERT, and Llama-7B have made significant progress in the field of natural language processing. These models have strong pattern recognition and reasoning capabilities and can perform well in a variety of natural language tasks. Some studies have tried to apply large models to time series prediction by converting time series data into text format and using large models for prediction. However, directly converting time series data into text may lose important time series features, and existing methods still have shortcomings in modal alignment and interpretability of prediction results. Therefore, existing time series prediction technologies face multiple challenges, especially lack of general application and adaptability. Specifically, traditional models can only target specific scenarios and are difficult to adaptively adjust according to different types of data characteristics. This leads to less than ideal prediction accuracy in a variety of application scenarios.
[0004] In view of this, how to provide a time prediction method that can significantly improve the adaptability, versatility and prediction accuracy of the model has become a technical problem that needs to be solved urgently. Summary of the invention
[0005] Embodiments of the present application provide a time prediction system based on adaptive prefix generation, a time prediction method based on adaptive prefix generation, a time prediction device based on adaptive prefix generation, an electronic setting and a computer storage medium, which are used to solve the problems of low adaptability, low versatility and low prediction accuracy of current time prediction models.
[0006] In a first aspect of an embodiment of the present application, a time prediction system based on adaptive prefix generation is provided, including:
[0007] Input data processing module, used to preprocess time series data;
[0008] A data conversion module, used to convert each data module of the time series data into a text prototype representation that meets the format requirements of the target language model;
[0009] The adaptive prefix generation module is used to understand and predict the time series through the adaptive prefix generation mechanism and generate the input time series prefix;
[0010] A pre-trained language module, used for processing and predicting the input time series prefix to generate output features;
[0011] The output projection module is used to convert the output features generated by the pre-trained language module into a time series prediction result.
[0012] In a second aspect of an embodiment of the present application, a time prediction method based on adaptive prefix generation is provided, including:
[0013] Preprocessing the original time series data to generate a data module, and converting the data module into a text prototype representation that meets the format requirements of the target language model;
[0014] Through the adaptive prefix generation mechanism, the time series is understood and predicted, and the input time series prefix is generated;
[0015] Inputting the input time series prefix into the pre-trained language module for processing and prediction to generate output features;
[0016] The output features generated by the pre-trained language module are converted into time series prediction results.
[0017] In a third aspect of an embodiment of the present application, a time prediction device based on adaptive prefix generation is provided, including:
[0018] A format conversion module is configured to preprocess the original time series data, generate a data module, and convert the data module into a text prototype representation that meets the format requirements of the target language model;
[0019] The prefix generation module is configured to understand and predict the time series and generate a prefix of the input time series through an adaptive prefix generation mechanism;
[0020] A feature generation module is configured to input the input time series prefix into the pre-trained language module for processing and prediction, and generate output features;
[0021] The result conversion module is configured to convert the output features generated by the pre-trained language module into time series prediction results.
[0022] According to a fourth aspect of an embodiment of the present application, a computing device is provided, including:
[0023] Memory and processor;
[0024] The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the above-mentioned time prediction method based on adaptive prefix generation are implemented.
[0025] According to a fifth aspect of an embodiment of the present application, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the instructions are executed by a processor, the steps of the above-mentioned time prediction method based on adaptive prefix generation are implemented.
[0026] The present application provides a time prediction system based on adaptive prefix generation, comprising: an input data processing module for preprocessing time series data; a data conversion module for converting each data module of the time series data into a text prototype representation that meets the format requirements of a target language model; an adaptive prefix generation module for understanding and predicting the time series through an adaptive prefix generation mechanism to generate an input time series prefix; a pre-trained language module for processing and predicting the input time series prefix to generate output features; and an output projection module for converting the output features generated by the pre-trained language module into a time series prediction result.
[0027] The time prediction system based on adaptive prefix generation provided by the embodiment of the present application is applied, and the performance of the time series prediction model is significantly improved through data conversion and adaptive prefix generation mechanism. The data conversion module converts the time series data into a text prototype that can be processed by the language model, which solves the problem of aligning the time series with the natural language modality, and enables the pre-trained language model to predict the time series more accurately. The adaptive prefix generation mechanism enhances the model's understanding and prediction ability of time series data by dynamically generating prefixes that match specific tasks and contextual information, so that the model can be adaptively adjusted according to different task requirements, and provide more targeted and explanatory prediction results. The combination of these two mechanisms not only improves the prediction accuracy and robustness of the model, but also enhances the versatility and adaptability of the model.
[0028] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, which can be implemented in accordance with the contents of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:
[0030] Figure 1 A schematic diagram of the structure of a time prediction system based on adaptive prefix generation provided in an embodiment of the present application;
[0031] Figure 2 A schematic diagram of a flow chart of a time prediction method based on adaptive prefix generation provided in an embodiment of the present application;
[0032] Figure 3 A schematic diagram of the structure of a time prediction device based on adaptive prefix generation provided in an embodiment of the present application;
[0033] Figure 4 A structural block diagram of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0034] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.
[0035] The current time prediction models include a track prediction method based on Kalman filtering and a track prediction method based on LSTM.
[0036] Specifically, the structure and relationship of the track prediction method based on Kalman filtering include: the Kalman filter is composed of a predictor and an updater. The predictor makes predictions based on the system model and prior estimates, and the updater uses measurement data to correct the prediction results. The two work alternately to achieve real-time estimation of the target state.
[0037] The method steps include: To establish a state space model, it is first necessary to determine the state equation and observation equation of the system. The state equation describes the dynamic changes of the system, and the observation equation describes the relationship between the observed data and the state. In this process, we need to define the state variables (such as position, velocity, etc.), as well as the covariance matrix of process noise and observation noise. During initialization, it is necessary to set the initial state estimate and the initial estimate error covariance matrix. Next, in the prediction step, according to the state equation of the system, the state estimate at the previous moment is used to predict the state at the current moment, and the prior state estimate and the prior estimate error covariance matrix are calculated. In the update step, the observation data at the current moment is obtained, and the observation error, that is, the difference between the actual observation value and the prior observation value, is calculated. Then, the Kalman gain is calculated, and the state estimate and the estimate error covariance matrix are updated. The whole process realizes real-time update of the system state estimate by iteratively executing the prediction and update steps.
[0038] The structural composition and relationship of the LSTM-based track prediction method include: the LSTM network is composed of multiple LSTM units, each unit contains an input gate, a forget gate and an output gate, and the transmission and forgetting of information are controlled through the gating mechanism. The input layer accepts time series data, the hidden layer is composed of multiple LSTM units, and the output layer generates prediction results.
[0039] The method steps include: in the data preprocessing stage, the original track data is cleaned, normalized and segmented to form time series data. The cleaning process includes removing noise and outliers, and the normalization process scales the data to a uniform range. In the model building stage, an LSTM model is established, including an input layer, multiple LSTM layers and an output layer. The input layer receives the preprocessed time series data, each LSTM layer captures the dependencies in the data through memory units and gate mechanisms, and the output layer generates prediction results for future time points. In the model training stage, the preprocessed track data is used to train the LSTM model, the model parameters are adjusted to minimize the prediction error, and the back propagation algorithm is used to update the parameters to optimize the model performance. Finally, in the track prediction stage, the track data collected in real time is input into the trained LSTM model to perform track prediction, and the predicted track data for a period of time in the future is output, including position information and speed.
[0040] Existing time series prediction models are usually designed for specific tasks, lack versatility and adaptability, and are difficult to be widely used in different application scenarios. Existing models such as Kalman filtering, although they perform well in specific linear time series prediction tasks, are less effective when processing data with complex patterns and nonlinear relationships. On the other hand, although machine learning models such as LSTM have improved in capturing nonlinear relationships, they require a large amount of task-specific data for training, and their model structure is complex and difficult to tune.
[0041] In addition, these models need to be redesigned and trained for each new task in practical applications, resulting in high migration costs and low efficiency between different tasks. For example, deep learning models such as LSTM may need to be re-tuned or even rebuilt in different application scenarios, which consumes a lot of time and computing resources. Therefore, it is difficult for these models to quickly adapt to new time series prediction tasks, limiting their breadth and flexibility in practical applications.
[0042] Therefore, the present application solves the deficiencies of the existing time series prediction model in terms of versatility and adaptability by introducing an adaptive prefix generation mechanism. By utilizing the contextual information and task instructions provided by natural language prompts, the time series data is converted into a format that can be understood by the language model, thereby enhancing the versatility and adaptability of the model and reducing the need for redesign and training. Through improvements, the present invention significantly improves the practicality and application prospects of the time series prediction model.
[0043] See also Figure 1 , Figure 1 A schematic diagram of the structure of a time prediction system based on adaptive prefix generation provided in an embodiment of the present application. Figure 1 As shown, the time prediction system based on adaptive prefix generation includes an input data processing module 101, a data conversion module 102, an adaptive prefix generation module 103, a pre-trained language module 104 and an output projection module 105.
[0044] The input data processing module 101 is used to pre-process the time series data.
[0045] The data conversion module 102 is used to convert each data module of the time series data into a text prototype representation that meets the format requirements of the target language model.
[0046] The adaptive prefix generation module 103 is used to understand and predict the time series through an adaptive prefix generation mechanism to generate a prefix of the input time series.
[0047] The pre-trained language module 104 is used to process and predict the input time series prefix to generate output features.
[0048] The output projection module 105 is used to convert the output features generated by the pre-trained language module into a time series prediction result.
[0049] Specifically, the input data processing module 101 includes a normalization processing unit 1011 and a data segmentation unit 1012, wherein the normalization processing unit 1011 is used to normalize the time series data according to the original time series data, the mean and standard deviation of the time series; the data segmentation unit 1012 is used to divide the normalized time series to obtain data modules.
[0050] In practical applications, the specific implementation steps of normalization processing are as follows: Let the original time series data be X = {x1, x2, ..., x T}, after normalization, we get X'={x1',x2',......x T '},in:
[0051]
[0052] Where μ is the mean of the time series and σ is the standard deviation.
[0053] The specific implementation steps of data segmentation are: divide the normalized time series X' into multiple patch modules of length L, each patch is p i , every L / 2 interval, that is:
[0054] P i ={x i ',x i+1 ',......,x i+L-1 '}
[0055] Here, i is the starting time step.
[0056] The data conversion module 102 includes a linear transformation unit 1021 and a cross-attention mechanism unit 1022, wherein the linear transformation unit 1021 is used to combine and project the text prototypes into a small sample text set through a fully connected layer; the cross-attention mechanism unit 1022 is used to align the time series features and the word embedding of the language model through a cross-attention mechanism, calculate the attention weight, and calculate the weighted feature representation based on the attention weight.
[0057] In practical applications, the specific implementation steps of linear transformation are as follows: text prototype set E i Projected to T through the fully connected layer i , that is: T i =R*E i .
[0058] Among them, R is the mapping matrix, E i is a set of text prototypes, T i It is a small sample text collection.
[0059] The specific implementation steps of the cross attention mechanism are: align the time series features T through the cross attention mechanism i Word embedding T with language model i , (where the time series input feature is Q, and the words of the language model are K and V).
[0060] Calculate the attention weight a:
[0061] Then, the weighted feature representation F is calculated: F = a*V.
[0062] The adaptive prefix generation module 103 includes an adaptive prefix generation unit 1031 and a prefix generation unit 1032, wherein the adaptive prefix generation unit 1031 is used to input time series task information and context information into a large model and output an adaptive prefix prompt; the prefix generation unit 1032 is used to use the time series features after patch reprogramming as a prefix, combined with the adaptive prefix prompt, to generate an input time series prefix.
[0063] In practical applications, the specific implementation steps of the adaptive prefix generation module are as follows: Generate adaptive prefix hints through a large model, the specific process is as follows: P = LM (E A , E B ).
[0064] Among them, E A is the time series task information, E B The large model LM generates a prefix P that matches the input task and context information. P is an adaptive prefix that is dynamically generated based on the specific task and provides context information and task instructions.
[0065] The specific implementation steps of prefix generation are: using the time series feature F after patch reprogramming as a prefix, combined with the natural language prompt P, to form an input sequence S: S = {F, P}.
[0066] The pre-trained language module 104 includes a model selection unit 1041 and a freezing parameter unit 1042, wherein the model selection unit 1041 is used to select a target pre-trained language model according to its level requirements; the freezing parameter unit 1042 is used to freeze the parameters of the target pre-trained language model during the model training process so that the target pre-trained language model processes the input data.
[0067] In practical applications, the specific implementation steps of model selection are: select a suitable pre-trained language model, such as BERT or Llama, and select the appropriate model size and parameters based on your own machine level requirements.
[0068] The specific implementation steps of freezing parameters are: during the model training process, freeze the parameters of the large model, do not train the parameters of the large model, only process and optimize the input data, and reduce computing resources and training time.
[0069] The language model processes the input sequence S and generates output features O: O = LM(S).
[0070] The output projection module 105 includes a flattening processing unit 1051, a linear projection unit 1052 and a feedback unit 1053 of an adaptive module, wherein the flattening processing unit 1051 is used to flatten the output features generated by the pre-trained language module into a one-dimensional vector; the linear projection unit 1052 is used to project the one-dimensional vector to a target prediction value space through a linear layer to generate a time series prediction result; the feedback unit 1053 of the adaptive module is used to input the time series prediction result and the time series prefix into a pre-trained language model for optimization, and output the optimization result.
[0071] In practical applications, the specific implementation steps of the flattening process are: flattening the output feature O of the language model into a one-dimensional vector O': O'=flatten(O).
[0072] The specific implementation steps of linear projection are: project the flattened feature vector O' to the target prediction value space Y through the linear layer to generate the final time series prediction result: Y = W y O'.
[0073] Among them, W y is the projection matrix, and Y is the prediction result.
[0074] The specific implementation steps of the feedback of the adaptive module are: input the prediction results and prefixes into the big model, prompting the big model that the prefix needs to be optimized to further improve the matching of the prefix.
[0075] The time prediction system based on adaptive prefix generation provided by the embodiment of the present application is applied, and the performance of the time series prediction model is significantly improved through data conversion and adaptive prefix generation mechanism. The data conversion module converts the time series data into a text prototype that can be processed by the language model, which solves the problem of aligning the time series with the natural language modality, and enables the pre-trained language model to predict the time series more accurately. The adaptive prefix generation mechanism enhances the model's understanding and prediction ability of time series data by dynamically generating prefixes that match specific tasks and contextual information, so that the model can be adaptively adjusted according to different task requirements, and provide more targeted and explanatory prediction results. The combination of these two mechanisms not only improves the prediction accuracy and robustness of the model, but also enhances the versatility and adaptability of the model.
[0076] See also Figure 2 , Figure 2 A flow chart of a time prediction method based on adaptive prefix generation provided in an embodiment of the present application. Figure 2 As shown, the specific steps include the following steps.
[0077] Step S202: pre-processing the original time series data to generate a data module, and converting the data module into a text prototype representation that meets the format requirements of the target language model.
[0078] Step S204: The time series is understood and predicted through an adaptive prefix generation mechanism to generate an input time series prefix.
[0079] Step S206: inputting the input time series prefix into a pre-trained language module for processing and prediction to generate output features.
[0080] Step S208: converting the output features generated by the pre-trained language module into time series prediction results.
[0081] The time prediction method based on adaptive prefix generation provided by the embodiment of the present application is applied, and the performance of the time series prediction model is significantly improved through data conversion and adaptive prefix generation mechanism. The data conversion module converts the time series data into a text prototype that can be processed by the language model, which solves the problem of aligning the time series with the natural language modality, and enables the pre-trained language model to predict the time series more accurately. The adaptive prefix generation mechanism enhances the model's understanding and prediction ability of time series data by dynamically generating prefixes that match specific tasks and contextual information, so that the model can be adaptively adjusted according to different task requirements, and provide more targeted and explanatory prediction results. The combination of these two mechanisms not only improves the prediction accuracy and robustness of the model, but also enhances the versatility and adaptability of the model.
[0082] The specific implementation method can refer to the above description of the time prediction system based on adaptive prefix generation, and the embodiments of the present application will not be described in detail.
[0083] Corresponding to the above method embodiment, this specification also provides a time prediction device embodiment based on adaptive prefix generation, Figure 3 A schematic diagram of the structure of a time prediction device based on adaptive prefix generation provided in an embodiment of the present application. Figure 3 As shown, the device comprises:
[0084] The format conversion module 302 is configured to pre-process the original time series data, generate a data module, and convert the data module into a text prototype representation that meets the format requirements of the target language model;
[0085] The prefix generation module 304 is configured to understand and predict the time series through an adaptive prefix generation mechanism to generate a prefix for the input time series;
[0086] The feature generation module 306 is configured to input the input time series prefix into the pre-trained language module for processing and prediction, and generate output features;
[0087] The result conversion module 308 is configured to convert the output features generated by the pre-trained language module into time series prediction results.
[0088] The time prediction device based on adaptive prefix generation provided by the embodiment of the present application significantly improves the performance of the time series prediction model through data conversion and adaptive prefix generation mechanism. The data conversion module converts the time series data into a text prototype that can be processed by the language model, which solves the problem of aligning the time series with the natural language modality, and enables the pre-trained language model to predict the time series more accurately. The adaptive prefix generation mechanism enhances the model's understanding and prediction capabilities of time series data by dynamically generating prefixes that match specific tasks and contextual information, so that the model can be adaptively adjusted according to different task requirements, and provide more targeted and explanatory prediction results. The combination of these two mechanisms not only improves the prediction accuracy and robustness of the model, but also enhances the versatility and adaptability of the model.
[0089] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the time prediction device based on adaptive prefix generation, since it is basically similar to the embodiment of the time prediction method based on adaptive prefix generation, the description is relatively simple, and the relevant parts can be referred to the partial description of the embodiment of the time prediction method based on adaptive prefix generation.
[0090] Figure 4 This is a block diagram of a computing device provided in an embodiment of the present application. The components of the computing device 400 include but are not limited to a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and the database 450 is used to store data.
[0091] The computing device 400 also includes an access device 440, which enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 440 may include one or more of any type of network interface (e.g., a network interface card (NIC)) of wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a world-wide interoperability for microwave access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, and a near field communication (NFC).
[0092] In one embodiment of the present specification, the above components of the computing device 400 and Figure 4 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 4 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.
[0093] The computing device 400 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 400 may also be a mobile or stationary server.
[0094] The processor 420 is used to execute the following computer executable instructions, which, when executed by the processor, implement the steps of the above-mentioned time prediction method based on adaptive prefix generation.
[0095] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the computing device embodiment, since it is basically similar to the time prediction method embodiment based on adaptive prefix generation, the description is relatively simple, and the relevant parts can be referred to the partial description of the time prediction method embodiment based on adaptive prefix generation.
[0096] An embodiment of the present specification further provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned time prediction method based on adaptive prefix generation.
[0097] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the computer-readable storage medium embodiment, since it is basically similar to the time prediction method embodiment based on adaptive prefix generation, the description is relatively simple, and the relevant parts can be referred to the partial description of the time prediction method embodiment based on adaptive prefix generation.
[0098] An embodiment of the present specification further provides a computer program, wherein when the computer program is executed in a computer, the computer is instructed to execute the steps of the above-mentioned time prediction method based on adaptive prefix generation.
[0099] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the computer program embodiment, since it is basically similar to the time prediction method embodiment based on adaptive prefix generation, the description is relatively simple, and the relevant parts can be referred to the partial description of the time prediction method embodiment based on adaptive prefix generation.
[0100] The above is a description of a specific embodiment of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0101] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.
[0102] It should be noted that the above is a description of a specific embodiment of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of the present specification.
[0103] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0104] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can well understand and use this specification. This specification is only limited by the claims and their full scope and equivalents.
Claims
1. A time prediction system based on adaptive prefix generation, characterized in that: include: Input data processing module, used to preprocess time series data; A data conversion module, used to convert each data module of the time series data into a text prototype representation that meets the format requirements of the target language model; The adaptive prefix generation module is used to understand and predict the time series through the adaptive prefix generation mechanism and generate the input time series prefix; A pre-trained language module, used for processing and predicting the input time series prefix to generate output features; The output projection module is used to convert the output features generated by the pre-trained language module into a time series prediction result.
2. The time prediction system based on adaptive prefix generation according to claim 1, characterized in that: The input data processing module includes a normalization processing unit and a data segmentation unit, wherein: The normalization processing unit is used to perform normalization processing on the time series data according to the original time series data, the mean value and the standard deviation of the time series; The data segmentation unit is used to divide the normalized time series to obtain data modules.
3. The time prediction system based on adaptive prefix generation according to claim 1, characterized in that: The data conversion module includes a linear transformation unit and a cross attention mechanism unit, wherein: The linear transformation unit is used to project the text prototype into a small sample text set through a fully connected layer; The cross attention mechanism unit is used to align the time series features and the word embedding of the language model through the cross attention mechanism, calculate the attention weight, and calculate the weighted feature representation based on the attention weight.
4. The time prediction system based on adaptive prefix generation according to claim 1, characterized in that: The adaptive prefix generation module includes an adaptive prefix generation unit and a prefix generation unit, wherein: The adaptive prefix generation unit is used to input the time series task information and context information into the large model and output an adaptive prefix prompt; The prefix generation unit is used to use the time series feature after patch reprogramming as a prefix, combined with the adaptive prefix prompt, to generate an input time series prefix.
5. The time prediction system based on adaptive prefix generation according to claim 1, characterized in that: The pre-trained language module includes a model selection unit and a frozen parameter unit, wherein: The model selection unit is used to select a target pre-trained language model according to the level requirement; The freezing parameter unit is used to freeze the parameters of the target pre-trained language model during the model training process so that the target pre-trained language model processes the input data.
6. The time prediction system based on adaptive prefix generation according to claim 1, characterized in that: The output projection module includes a flattening processing unit, a linear projection unit and a feedback unit of an adaptive module, wherein: The flattening processing unit is used to flatten the output features generated by the pre-trained language module into a one-dimensional vector; The linear projection unit is used to project the one-dimensional vector to the target prediction value space through a linear layer to generate a time series prediction result; The feedback unit of the adaptive module is used to input the time series prediction result and the time series prefix into the pre-trained language model for optimization and output the optimization result.
7. A time prediction method based on adaptive prefix generation, characterized in that: include: Preprocessing the original time series data to generate a data module, and converting the data module into a text prototype representation that meets the format requirements of the target language model; Through the adaptive prefix generation mechanism, the time series is understood and predicted, and the input time series prefix is generated; Inputting the input time series prefix into the pre-trained language module for processing and prediction to generate output features; The output features generated by the pre-trained language module are converted into time series prediction results.
8. A time prediction device based on adaptive prefix generation, characterized in that: include: A format conversion module is configured to preprocess the original time series data, generate a data module, and convert the data module into a text prototype representation that meets the format requirements of the target language model; The prefix generation module is configured to understand and predict the time series and generate a prefix of the input time series through an adaptive prefix generation mechanism; A feature generation module is configured to input the input time series prefix into the pre-trained language module for processing and prediction, and generate output features; The result conversion module is configured to convert the output features generated by the pre-trained language module into time series prediction results.
9. A computing device comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method described in any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an implementation program for information transmission, and when the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.