Time sequence prediction method, time sequence prediction device and electronic equipment

Abstract

The invention is suitable for the technical field of computers, and provides a time sequence prediction method, a time sequence prediction device and electronic equipment. The method comprises the following steps: acquiring historical time sequence data of a target object; k target time sequence data samples matched with the historical time sequence data are determined from at least two time sequence data samples, and K is an integer larger than zero; determining at least one target expert model matched with the historical time sequence data based on expert index sets corresponding to the K target time sequence data samples; the expert index set comprises a corresponding relation between M expert models and performance indexes, the performance indexes represent the performance of the corresponding expert models when the corresponding time series data samples are subjected to time series prediction, and M is an integer greater than 1; and predicting target time sequence data of the target object in a target time period after the current moment based on the historical time sequence data and each target expert model. According to the invention, the time sequence prediction accuracy can be improved.

Description

Technical Field

[0001] The present application belongs to the field of computer technology, and in particular, relates to a time series prediction method, a time series prediction device, and an electronic device. Background Art

[0002] Time series data is used to describe how things or phenomena change over time. This type of data reflects the state or degree of change of a thing or phenomenon over time. Time series forecasting refers to the forecasting task for time series data, that is, predicting time series data for a period of time in the future based on known historical time series data.

[0003] Currently, there is a solution for time series prediction through a time series prediction model, but the time series prediction accuracy of this solution is low. Summary of the invention

[0004] The embodiments of the present application provide a time series prediction method, a time series prediction device, and an electronic device, which can improve the accuracy of time series prediction.

[0005] In a first aspect, an embodiment of the present application provides a time series prediction method, comprising: Obtain historical time series data of the target object; Determine, from at least two time series data samples, K target time series data samples that match the historical time series data, where K is an integer greater than zero; Based on the expert indicator sets corresponding to the K target time series data samples, at least one target expert model matching the historical time series data is determined; the expert indicator set includes: a correspondence between M expert models and performance indicators, the performance indicators characterizing the performance of the corresponding expert model when performing time series prediction on the corresponding time series data samples, where M is an integer greater than 1; Based on the historical time series data and each target expert model, the target time series data of the target object in the target time period after the current moment is predicted.

[0006] In an embodiment of the present application, by acquiring historical time series data of the target object, K target time series data samples that match the historical time series data can be determined from at least two time series data samples. Since the K target time series data samples match the historical time series data, and the performance indicators of the expert models in the expert indicator set characterize the performance of the corresponding expert model when performing time series prediction on the corresponding time series data samples, at least one target expert model that matches the historical time series data can be determined based on the expert indicator sets corresponding to the K target time series data samples. Based on the historical time series data and each target expert model, the target time series data of the target object in the target time period after the current moment can be predicted more accurately, thereby improving the accuracy of time series prediction.

[0007] In a second aspect, an embodiment of the present application provides a time series prediction device, including: A data acquisition module is used to acquire historical time series data of a target object; A sample determination module, configured to determine K target time series data samples matching the historical time series data from at least two time series data samples, where K is an integer greater than zero; A target determination module, configured to determine at least one target expert model matching the historical time series data based on the expert indicator sets corresponding to the K target time series data samples; the expert indicator set includes: a correspondence between M expert models and performance indicators, wherein the performance indicators represent the performance of the corresponding expert model when performing time series prediction on the corresponding time series data samples, and M is an integer greater than 1; The target prediction module is used to predict the target time series data of the target object in a target time period after the current moment based on the historical time series data and each target expert model.

[0008] In a third aspect, an embodiment of the present application 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 electronic device implements the timing prediction method as described in the first aspect above.

[0009] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the time series prediction method as described in the first aspect above is implemented.

[0010] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, and when the computer program is executed, the time series prediction method described in the first aspect above is executed.

[0011] It can be understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0013] Figure 1 is a schematic diagram of the architecture of a time series prediction system provided in an embodiment of the present application; Figure 2It is a flowchart of a time series prediction method provided in an embodiment of the present application; Figure 3 It is a framework diagram of the time series prediction method provided in the embodiment of the present application; Figure 4 It is a flowchart of the multi-expert model pre-training stage provided in an embodiment of the present application; Figure 5 It is a flowchart of the retrieval enhancement reasoning stage provided in an embodiment of the present application; Figure 6 is a schematic diagram of the structure of a timing prediction device provided in an embodiment of the present application; Figure 7 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0014] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0015] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0016] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0017] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0018] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0019] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0020] The timing prediction method provided in the embodiments of the present application can be applied to electronic devices such as mobile phones, tablet computers, laptop computers, servers, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), etc. The embodiments of the present application do not impose any restrictions on the specific types of electronic devices.

[0021] For example, Figure 1 Schematic diagram of the architecture of the time series prediction system provided by the embodiment of the present application is shown. Figure 1 As shown, the system includes a terminal 101 and a server 102. A communication connection can be established between the terminal 101 and the server 102 to achieve data transmission between the terminal 101 and the server 102.

[0022] The terminal 101 can obtain the historical time series data of the target object and send the historical time series data to the server 102. After receiving the historical time series data, the server 102 can perform time series prediction on the target object through the time series prediction scheme provided by the present application. Optionally, after predicting the target time series data of the target object in the target time period after the current moment, the server 102 can send the target time series data to the terminal 101, and the terminal 101 displays the target time series data so that the user can understand the target time series data.

[0023] See also Figure 2 , Figure 2 A flow chart of a time series prediction method provided in an embodiment of the present application is shown. As an example but not a limitation, the method is applied to an electronic device, and the method includes the following steps: Step 201, obtaining historical time series data of the target object.

[0024] The target object may be any object that needs to be time series predicted, and the present application does not limit the specific type of the target object.

[0025] The above-mentioned historical time series data may refer to the time series data of the target object in a specified time period. Optionally, the above-mentioned specified time period can be set according to the scenario requirements. As an example and not a limitation, the moments within the above-mentioned specified time period include moments earlier than the current moment (i.e., moments before the current moment), and may also include the current moment and moments earlier than the current moment. Since the closer the time, the more similar the corresponding data may be, the last moment of the specified time period can be the current moment. On this basis, the time series data of the target object after the current moment can be more accurately predicted based on the historical time series data of the target object.

[0026] Time series data can also be called time series data, which usually contains timestamps and attribute values, and is usually in the form of (t0, v0), (t1, v1),…, (tn, vn).

[0027] The timestamp t in this application (i.e., timestamps such as t0, t1, ..., tn) is a timestamp that changes at equal intervals, v (i.e., attribute values ​​such as v0, v1, ..., vn) is an attribute value that changes over time, and the dimension of the attribute value v is ≥ 1. For example, if the time series data is the hourly electricity consumption of a household, then the dimension of the corresponding attribute value v is 1; if the time series data is the daily turnover of a store, then the dimension of the corresponding attribute value v is 1; if the time series data is the annual sales and net profit of a company, then the dimension of the corresponding attribute value v is 2; if the time series data is the number of cars, trucks, and engineering vehicles passing through an intersection every hour, then the dimension of the corresponding attribute value v is 3. Historical time series data can be obtained from a terminal or a sensor corresponding to the terminal. For example, daily temperature, daily traffic volume, and vehicle speed are obtained through temperature sensors and speed sensors. For example, daily business flow data is obtained from a store cash register terminal.

[0028] Step 202: Determine K target time series data samples that match the historical time series data from at least two time series data samples.

[0029] Wherein, K is an integer greater than zero. The above-mentioned at least two time series data samples are pre-constructed time series data, which have the same attributes and duration as the historical time series data. The objects targeted by the above-mentioned two time series data samples may be the target object or may not be the target object, and the objects targeted by different time series data samples may be the same or different, which is not limited in this application.

[0030] As an example but not a limitation, there are three time series data samples. The historical time series data is the hourly electricity consumption of household A in the previous two days. Its attribute is hourly electricity consumption and the duration is 48 hours. Then these three time series data samples can be the hourly electricity consumption of household A in three different 48-hour periods, or they can be the hourly electricity consumption of household A in a certain 48 hours, the hourly electricity consumption of household B in a certain 48 hours, and the hourly electricity consumption of household C in a certain 48 hours.

[0031] In a possible implementation, the electronic device may determine a target similarity between each time series data sample and the historical time series data; and determine the time series data samples ranked in the top K positions in terms of target similarity as target time series data samples.

[0032] For any time series data sample, the target similarity between the time series data sample and the historical time series data is an indicator for measuring the similarity between the time series data sample and the historical time series data. The larger the value, the higher the similarity between the time series data sample and the historical time series data, and the higher the matching degree. On this basis, by using the time series data sample with the target similarity ranked in the top K as the target time series data sample, the target time series data sample that is more matched with the historical time series data can be screened out from at least two time series data samples. Optionally, the specific value of K can be set according to the scenario requirements, and this application does not limit this.

[0033] In one embodiment, the electronic device may determine the target similarity between each time series data sample and the historical time series data in the following four ways.

[0034] Method 1: Based on historical time series data, determine the target similarity between each time series data sample and the historical time series data.

[0035] For the above-mentioned method 1, the electronic device can calculate the first similarity between each time series data sample and the historical time series data based on each time series data sample and the historical time series data, and determine the first similarity between each time series data sample and the historical time series data as the target similarity between each time series data sample and the historical time series data.

[0036] As an example but not limitation, for time series data sample A, based on time series data sample A and historical time series data, a first similarity between time series data sample A and historical time series data can be calculated, and this first similarity is the target similarity between time series data sample A and historical time series data.

[0037] In a second method, the electronic device may construct a prompt for the historical time series data to obtain a first prompt word; based on the first prompt word, determine a target similarity between each time series data sample and the historical time series data.

[0038] For the above-mentioned method 2, the electronic device can calculate the second similarity between each time series data sample and the historical time series data based on the second prompt word of each time series data sample and the first prompt word of the historical time series data, and determine the second similarity between each time series data sample and the historical time series data as the target similarity between each time series data sample and the historical time series data.

[0039] As an example but not limitation, for time series data sample A, the similarity between the second prompt word of time series data sample A and the first prompt word of historical time series data is calculated (in order to distinguish it from the first similarity, the similarity of the prompt word can be referred to as the second similarity), and the second similarity is the target similarity between time series data sample A and historical time series data.

[0040] In a third method, the electronic device may construct a prompt for the historical time series data to obtain a first prompt word; based on the historical time series data and the first prompt word, determine a target similarity between each time series data sample and the historical time series data.

[0041] For the above-mentioned method three, the electronic device can calculate the first similarity between each time series data sample and the historical time series data based on each time series data sample and the historical time series data; calculate the second similarity between each time series data sample and the historical time series data based on the second prompt word of each time series data sample and the first prompt word of the historical time series data; and determine the target similarity between each time series data sample and the historical time series data based on the first and second similarities between each time series data sample and the historical time series data.

[0042] For any time series data sample, the electronic device can add the first similarity and the second similarity between the time series data sample and the historical time series data to obtain a target similarity; or determine the average of the first similarity and the second similarity as the target similarity; or pre-set the weight of the first similarity and the weight of the second similarity, the weight of the first similarity represents the importance of the first similarity relative to the target similarity, and the weight of the second similarity represents the importance of the second similarity relative to the target similarity, based on the weight of the first similarity and the weight of the second similarity, perform weighted summation of the first similarity and the second similarity to obtain the target similarity.

[0043] It should be noted that the present application does not limit the measurement method of the first similarity and the second similarity. For example, the measurement method may be any method such as Euclidean distance, cosine similarity, etc. For the first prompt word and the second prompt word, the electronic device may construct a prompt for the historical time series data based on the prompt template to obtain a corresponding prompt word (i.e., the first prompt word), and construct a prompt for each time series data sample based on the prompt template to obtain a corresponding prompt word (i.e., the second prompt word).

[0044] Among them, the prompt template can refer to a collection of multiple prompts (such as commands, scene descriptions, etc.), which can be used to stimulate the model to generate corresponding responses for different scenarios or tasks. Optionally, a prompt template can be constructed for time series data (such as historical time series data and time series data samples) according to the scenario or task requirements to record specific information of the time series data. For example, the prompt template includes but is not limited to the time information, attribute values, and event information within the corresponding time period of the time series data.

[0045] As an example but not limitation, the historical time series data includes (2023-9-30 20:00:00, 34), (2023-9-30 21:00:00, 53), (2023-9-30 22:00:00, 76), (2023-9-30 23:00:00, 25), and 34, 53, 76 and 25 in the historical time series data correspond to the electricity consumption in one hour. For example, 34 is the electricity consumption from 20:00:00 to 21:00:00. The first prompt word of the historical time series data can be "The start time of the input sequence is 2023-9-30 20:00:00, and the end time is 2023-9-30 23:00:00. This time period is before the National Day. The mean of the input sequence is 47, the variance is 382.5, the maximum value is 76, and the minimum value is 25. It rained lightly during this time period, and a traffic incident occurred."

[0046] The first prompt word contains semantic description information of the historical time series data, and the second prompt word contains semantic description information of the time series data sample. Therefore, based on the first prompt word and the second prompt word, it is possible to determine whether the historical time series data matches the time series data sample from a semantic perspective, thereby being able to more accurately find the time series data sample that matches the historical time series data. For example, for the above-mentioned method 2, when the time series data sample is relatively similar to the historical time series data, the two relatively similar time series data can be distinguished semantically based on the first prompt word and the second prompt word, thereby accurately obtaining the time series data sample that is relatively matching the historical time series data.

[0047] It should be noted that the prompt template for constructing the first prompt word and the second prompt word is the same template, on this basis, it can be ensured that the target time series data sample is matched from at least two time series data samples.

[0048] Since the prompt words (i.e., the first prompt words and the second prompt words) are usually in text form, before calculating the second similarity, the first prompt word can be characterized to obtain a first text semantic representation, and the second prompt word can be characterized to obtain a second text semantic representation. On this basis, the above-mentioned calculation of the second similarity between each time series data sample and the historical time series data based on the second prompt word of each time series data sample and the first prompt word of the historical time series data may include: calculating the second similarity between each time series data sample and the historical time series data based on the second text semantic representation of the second prompt word of each time series data sample and the first text semantic representation of the first prompt word of the historical time series data. Optionally, the first prompt word and the second prompt word can be characterized using a text embedding model to convert the first prompt word and the second prompt word from text into a vector representation for efficient information retrieval.

[0049] Method four, inputting historical time series data into a text embedding model to obtain a vector representation of the historical time series data, and concatenating the vector representation of the historical time series data with the first text semantic representation to obtain a first concatenated vector; inputting each time series data sample into a text embedding model to obtain a vector representation of each time series data sample, and concatenating the vector representation of each time series data sample with the corresponding second text semantic representation to obtain a second concatenated vector; calculating the similarity between each second concatenated vector and the first concatenated vector, and the similarity is the target similarity between the corresponding time series data sample and the historical time series data.

[0050] The fourth method mentioned above can determine whether the historical time series data matches the time series data sample from the perspectives of time series data and semantics, so as to more accurately find the time series data sample that matches the historical time series data.

[0051] Step 203: Determine at least one target expert model that matches the historical time series data based on the expert indicator sets corresponding to the K target time series data samples.

[0052] The expert indicator set includes: a correspondence between M expert models and performance indicators, the performance indicators represent the performance of the corresponding expert model when performing time series prediction on the corresponding time series data samples, and M is an integer greater than 1.

[0053] Since the K target time series data samples match the historical time series data of the target object, and the performance indicators characterize the performance of the corresponding expert model when performing time series prediction on the corresponding time series data samples, based on the performance indicators in the K expert indicator sets, a target expert model with better performance in time series prediction on historical time series data can be selected from the M expert models.

[0054] In one possible implementation, before step 202, the electronic device may first obtain at least two time series data samples and the true values ​​of at least two time series data samples, and on this basis, construct an expert indicator set corresponding to each time series data sample. Based on the expert indicator set, a target expert model that can more accurately perform time series prediction on the target object can be screened out from the M expert models.

[0055] For each time series data sample, the corresponding expert indicator set is constructed as follows: Input the time series data samples into M expert models respectively, and obtain the predicted values ​​of the time series data samples output by each of the M expert models; Based on the true value and predicted value of the time series data sample, determine the performance indicators of M expert models when performing time series prediction on the time series data sample; Based on the M expert models and the performance indicators of the M expert models when performing time series prediction on the time series data samples, an expert indicator set corresponding to the time series data samples is generated.

[0056] Optionally, the present application does not limit the measurement method of the performance index of the expert model. For example, the measurement method of the performance index of the expert model can be any method such as mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), etc.

[0057] For any time series data sample, the electronic device can establish a corresponding relationship between M expert models and performance indicators, thereby generating an expert indicator set corresponding to the time series data sample.

[0058] Before the electronic device inputs the time series data samples into the M expert models respectively, it can first train the M expert models, and the training process includes: The electronic device obtains a time series data set, and divides the time series data set into a training set, a validation set, and a test set according to a certain ratio; for all the time series data in the training set, the validation set, and the test set, a sample is constructed by sliding a window according to a certain input step length (i.e., the length of a specified time period) and an output step length (i.e., the length of a target time period); M time series prediction basic models are selected as expert models to form an expert library including M expert models; each expert model in the expert library is trained using samples in the training set and the validation set, and the trained M expert models are saved.

[0059] It should be understood that the training process of the above-mentioned M expert models can also be performed on the cloud platform. After the cloud platform trains the M expert models, the trained M expert models are saved. On this basis, before executing step 204, the electronic device can first obtain each target expert model from the cloud platform. After obtaining each target expert model from the cloud platform, the target object can be predicted in time series through each target expert model.

[0060] After obtaining the trained M expert models, the electronic device or cloud platform can select samples in the validation set and / or test set as time series data samples (for example, select samples in the validation set that are closer to the current moment as time series data samples), and generate a time series knowledge base for each time series data sample. The electronic device can perform time series prediction on the target object based on the time series knowledge base generated by itself or the time series knowledge base obtained from the cloud platform (for example, obtain at least two time series data samples from the time series knowledge base, obtain a second text semantic representation and an expert indicator set from the time series knowledge base, etc.).

[0061] When the first similarity between each time series data sample and the historical time series data is determined as the target similarity between each time series data sample and the historical time series data, the time series knowledge base may include each time series data sample and an expert indicator set corresponding to each time series data sample.

[0062] When the second similarity between each time series data sample and the historical time series data is determined as the target similarity between each time series data sample and the historical time series data, the above-mentioned time series knowledge base may include a second text semantic representation of the second prompt word of each time series data sample and an expert indicator set corresponding to each time series data sample.

[0063] When the target similarity between each time series data sample and the historical time series data is determined based on the first similarity and the second similarity between each time series data sample and the historical time series data, the above-mentioned time series knowledge base may include each time series data sample, a second text semantic representation of the second prompt word of each time series data sample, and an expert indicator set corresponding to each time series data sample.

[0064] In a possible implementation, the above step 203 may include: For any expert indicator set, N candidate expert models are determined from the expert indicator set; the N candidate expert models are expert models whose performance indicators are ranked in the top N positions, N is an integer greater than 0, and K target time series data samples correspond to K*N candidate expert models; Different candidate expert models among the K*N candidate expert models are determined as target expert models.

[0065] For any expert indicator set, the larger the performance indicator it contains, the better the performance of the corresponding expert model in time series prediction of the corresponding target time series data sample, and the target time series data sample matches the historical time series data. Therefore, the expert model with performance indicators ranked in the top N can better perform time series prediction on the historical time series data.

[0066] For each target time series data sample, N candidate expert models are selected from the corresponding expert indicator set. Then K target time series data samples correspond to K*N candidate expert models. There may be repeated candidate expert models among the K*N candidate expert models. In order to avoid using repeated candidate expert models for time series prediction of historical time series data, different candidate expert models among the K*N candidate expert models can be determined as the target expert model.

[0067] Step 204 , based on the historical time series data and each target expert model, predict the target time series data of the target object in the target time period after the current moment.

[0068] Optionally, the target time period may be set according to scenario requirements, which is not limited in this application.

[0069] Since each target expert model is an expert model selected from M different expert models and can perform time series prediction on historical time series data well, time series prediction can be performed on historical time series data based on each target expert model to obtain target time series data.

[0070] In a possible implementation, the above step 204 may include: Input the historical time series data into each target expert model to obtain the candidate time series data of the target object predicted by each target expert model in the target time period; Based on the weight of each target expert model, the candidate time series data predicted by each target expert model are weighted summed to obtain the target time series data.

[0071] In one embodiment, the weight of each target expert model may be set according to scenario requirements.

[0072] In another embodiment, on the basis of determining different candidate expert models among the K*N candidate expert models as target expert models, the weight of each target expert model is determined as follows: Based on the performance indicators corresponding to the K*N candidate expert models, determine at least one performance indicator corresponding to each target expert model; Determining a target performance indicator corresponding to each target expert model based on at least one performance indicator corresponding to each target expert model; Based on the target performance indicators corresponding to each target expert model, the weight of each target expert model is determined.

[0073] Since the K*N candidate expert models come from N expert indicator sets, there may be repeated candidate expert models among the K*N candidate expert models; the repeated candidate expert models come from different expert indicator sets, and each of them corresponds to a performance indicator in a different expert indicator set, so the target expert model corresponding to the repeated candidate expert model has at least two performance indicators; the non-repeated candidate expert models come from one expert indicator set, so the target expert model corresponding to the non-repeated candidate expert model has one performance indicator.

[0074] For a target expert model corresponding to a performance indicator, the performance indicator corresponding to the target expert model can be used as the target performance indicator of the target expert model; For a target expert model corresponding to at least two performance indicators, the target performance indicator of the target expert model can be determined based on the at least two performance indicators. Optionally, an average value of the at least two performance indicators corresponding to the target expert model can be calculated and the average value can be determined as the target performance indicator of the target expert model.

[0075] The electronic device may normalize the target performance indicators corresponding to each target expert model to obtain the weight of each target expert model. Optionally, the present application does not limit the normalization method of each target performance indicator. For example, a Softmax function may be used to normalize each target performance indicator.

[0076] In this embodiment, by acquiring the historical time series data of the target object, K target time series data samples that match the historical time series data can be determined from at least two time series data samples. Since the K target time series data samples match the historical time series data, and the performance indicators of the expert models in the expert indicator set characterize the performance of the corresponding expert model when performing time series prediction on the corresponding time series data samples, at least one target expert model that matches the historical time series data can be determined based on the expert indicator sets corresponding to the K target time series data samples. Based on the historical time series data and each target expert model, the target time series data of the target object in the target time period after the current moment can be predicted more accurately, thereby improving the accuracy of time series prediction.

[0077] Since the present application combines the above-mentioned M expert models to perform time series prediction, the above-mentioned M expert models can be understood as a time series Mixture of Experts (MoE) model.

[0078] The time series MoE model can refer to: in the time series prediction task, multiple expert models are combined, and for specific time series data, the corresponding expert model is activated through a certain strategy for reasoning, thereby achieving time series prediction.

[0079] Since the present application performs similarity retrieval from a time series knowledge base based on historical time series data and a first prompt word to retrieve target time series data samples that match the historical time series data, this retrieval method can be understood as retrieval enhancement in Retrieval-augmented Generation (RAG).

[0080] In the field of Natural Language Processing (NLP), RAG is a technology that uses information from private or proprietary data sources to supplement text generation. The present application can apply retrieval enhancement to time series prediction tasks, and use retrieval enhancement as an expert activation strategy in the above-mentioned time series MoE model. Therefore, a time series prediction method provided by the present application can also be referred to as a time series MoE prediction method based on retrieval enhancement, which is suitable for time series prediction scenarios. For example, it is suitable for scenarios such as traffic flow prediction in the transportation industry, water level prediction in the water conservancy industry, and electricity consumption prediction in the power industry. The method mainly includes a multi-expert model training stage and a retrieval enhancement reasoning stage.

[0081] like Figure 3 Shown is a framework diagram of the timing prediction method provided in an embodiment of the present application.

[0082] like Figure 3 As shown, the multi-expert model training phase mainly involves expert model training and time series knowledge base construction. (1) The M expert models may include large models (such as TimeGPT, Amazon's Chronos) and / or small models (such as Itransformer) in the field of time series prediction, and each expert model is trained separately based on the training set; (2) Based on the trained expert model and text embedding model, a time series knowledge base is constructed based on time series context samples including time series data samples, second prompt words, and true values. Constructing the time series knowledge base includes: inputting the second prompt word of the time series data sample into the text embedding model to obtain the second text semantic representation, inputting the time series data sample into the M expert models respectively to obtain the performance indicators of the M expert models, and constructing an expert indicator set; using the time series data sample and the second prompt word as the key, and the expert indicator set as the value, to construct the time series knowledge base.

[0083] like Figure 3As shown in the figure, the retrieval enhancement reasoning stage mainly performs retrieval routing and expert model activation. (1) Retrieval routing: Input the first prompt word of the historical time series data into the text embedding model to obtain the first text semantic representation; perform similarity retrieval from the time series knowledge base based on the first text semantic representation and the historical time series data, recall the keys ranked in the top K similarity, and obtain the expert indicator set corresponding to each key; (2) Expert model activation: Activate the expert model (i.e., candidate expert model) whose performance indicator corresponding to each key ranks in the top N, and obtain a total of K*N candidate expert models with repeated expert models; perform hybrid reasoning on the K*N candidate expert models according to the corresponding weights to achieve time series prediction.

[0084] like Figure 4 The figure is a flow chart of the multi-expert model pre-training stage provided in the embodiment of the present application. Figure 3 A detailed introduction to the training phase of the multi-expert model. Figure 4 The expert model training process shown includes: 1. Dataset division and sample construction: First, divide the data set into training set Dtrain, validation set Dval and test set Dtest in a ratio of 6:2:2; construct samples and true values ​​of samples according to the input step length input_len and the output step length pred_len sliding window for all time series data in the training set Dtrain, validation set Dval and test set Dtest, and finally obtain the samples and true values ​​of samples. Taking the i-th sample as an example, the sample can be represented as input_series_i and the true value can be represented as true_series_i.

[0085] 2. Time series prediction expert training: First, select M time series prediction base models as expert models to form an expert library; use the samples and true values ​​(input_series_i, true_series_i) in the training set Dtrain and the validation set Dval to train each expert model in the expert library and save the model file to obtain the trained expert library M1 ={expert_x}, x∈[1,M], where expert_x represents the xth trained expert model.

[0086] like Figure 4 The training process of constructing the temporal knowledge base shown includes: 1. Temporal context generation: Step 1.1, first select the last S samples in the validation set Dval as the sample set for generating temporal context samples, denoted as D1=(input_series_j, true_series_j), j∈[1,S]; Step 1.2, for the samples in D1 and the true value (input_series_j, true_series_j), construct prompts for the sample input_series_j according to the prompt template to obtain the corresponding prompt prompt_j; Step 1.3, repeat step 1.2, construct prompts for all samples in D1, and finally obtain the temporal context sample set D2 as ((input_series_j, prompt_j), true_series_j), j∈[1,S].

[0087] 2. Time series knowledge base construction: Step 2.1, input input_series_j in D2 into each expert model in M1 ={expert_x}, x∈[1,M], calculate the MSE (the smaller the better) between the predicted value pred_series_j and the true value true_series_j inferred by each expert model, and obtain the MSE indicator set of all expert models in M1 for input_series_j as {mse_x}, x∈[1,m]; Step 2.2, for input_series_j, form the corresponding expert indicator set EM_j=(expert_x, mse_x), x∈[1,M]; then sort the elements in the expert indicator set EM_j in ascending order according to MSE to obtain the ascending expert indicator set EM_sorted_j corresponding to input_series_j; Step 2.3, for input_series_j, use the text embedding model to represent its corresponding prompt_j, and obtain the second text semantic representation corresponding to input_series_j, recorded as text_emb_j; Step 2.4, for each sample in D2, repeat steps 2.1, 2.2 and 2.3, and finally obtain the ascending expert indicator set corresponding to each sample, and form a dictionary with the sample and the corresponding ascending expert indicator set to obtain the time series knowledge base TSKB={(input_series_j,text_emb_j): EM_sorted_j}, j ∈[1,S].

[0088] like Figure 5 The figure is a flow chart of the retrieval enhancement reasoning stage provided by the embodiment of the present application, which is a flowchart of Figure 3 A detailed description of the enhanced reasoning stage of retrieval is given in the following. Figure 5 The retrieval enhancement reasoning stage shown includes: 1. Retrieval routing: First, for the historical time series data input_series_r in the reasoning stage, construct a prompt for the historical time series data input_series_r according to the prompt template, obtain the corresponding first prompt word prompt_r, and form the historical time series context (input_series_r, prompt_r); according to step 2.3 of the time series knowledge base construction, obtain the first text semantic representation corresponding to input_series_r, recorded as text_emb_r; input_series_r, text_emb_r are respectively compared with the time series knowledge base TSKB={(input_series_j,text_emb_j): EM_sorted_j}, j ∈[1,num], that is, input_series_r is calculated with input_series_j in turn, and text_emb_r is calculated with text_emb_j in turn, and then the two similarities are added to get the K keys with the highest similarity to (input_series_r, text_emb_r); the corresponding values ​​of the above K keys in TSKB are taken out to obtain the K expert libraries corresponding to input_series_r, denoted as Mr={EM_sorted_b=(expert_c, mse_c)}, c∈[1,M], b∈[1,K].

[0089] 2. Activation and reasoning: First, activate the expert models arranged in N positions corresponding to the K expert libraries in Mr, and obtain Mr_activate = {EM_sorted_b = (expert_d, mse_d)}, d∈[1,N], b∈[1,K]; for the K*N expert models in Mr_activate (there are repeated expert models), calculate the MSE mean corresponding to each non-repeated expert model. For example, if there are V (V≤m) non-repetitive expert models in K*N, then calculate the MSE mean corresponding to each of the V expert models, and record the MSE mean corresponding to expert model v as mse_avg_v, and finally get (expert_v, mse_avg_v), v∈[1,V]; input_series_r into expert_v, v∈[1,V], and get V candidate time series data; calculate Softmax(mse_avg_v) to get the weights of the V expert models, and the sum of the weights is 1; multiply the V candidate time series data by the weights of the V expert models and add them together to get the final target time series data.

[0090] It should be understood that Figure 3 , Figure 4 and Figure 5Req1 in includes time series data sample 1 and the corresponding second text semantic representation, Req2 includes time series data sample 2 and the corresponding second text semantic representation, ReqN includes time series data sample N and the corresponding second text semantic representation, and ReqIn includes historical time series data and the corresponding first text semantic representation. Expert indicator set 1 is the expert indicator set corresponding to time series data sample 1, expert indicator set 2 is the expert indicator set corresponding to time series data sample 2, and expert indicator set S is the expert indicator set corresponding to time series data sample S.

[0091] It is worth noting that the above-mentioned multi-expert model training stage and the above-mentioned retrieval enhancement reasoning stage can be executed by the same device or by different devices, and this application does not limit this.

[0092] For example, the electronic device executes the above-mentioned multi-expert model training phase to obtain M trained expert models and a temporal knowledge base, saves the trained M expert models and the temporal knowledge base, and executes the above-mentioned retrieval enhancement reasoning phase based on the temporal knowledge base and the target expert model among the M expert models.

[0093] For example, the cloud platform executes the above-mentioned multi-expert model training phase to obtain M trained expert models and a time series knowledge base, and saves the trained M expert models and the time series knowledge base. The electronic device obtains the time series knowledge base and the target expert model among the M expert models from the cloud platform, and executes the above-mentioned retrieval enhanced reasoning phase based on the time series knowledge base and each target expert model.

[0094] Alternatively, the above-mentioned multi-expert model training phase and the above-mentioned retrieval enhancement reasoning phase are performed by different electronic devices. For example, the above-mentioned multi-expert model training phase is performed by electronic device 1 to obtain M trained expert models and a time series knowledge base, and the trained M expert models and the time series knowledge base are saved. Electronic device 2 obtains the time series knowledge base and the target expert model among the M expert models from electronic device 1, and performs the above-mentioned retrieval enhancement reasoning phase based on the time series knowledge base and each target expert model. It should be understood that the size of the serial number of each step in the above-mentioned embodiment does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0095] Corresponding to the time series prediction method described in the above embodiment, Figure 6 A schematic diagram of the structure of a timing prediction device provided in an embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0096] Reference Figure 6 , the device comprises: The data acquisition module 601 is used to acquire the historical time series data of the target object; A sample determination module 602 is used to determine K target time series data samples matching the historical time series data from at least two time series data samples, where K is an integer greater than zero; The target determination module 603 is used to determine at least one target expert model matching the historical time series data based on the expert indicator sets corresponding to the K target time series data samples; the expert indicator set includes: a correspondence between M expert models and performance indicators, the performance indicators characterizing the performance of the corresponding expert model when performing time series prediction on the corresponding time series data samples, and M is an integer greater than 1; The target prediction module 604 is used to predict the target time series data of the target object in the target time period after the current moment based on the historical time series data and each target expert model.

[0097] Optionally, the target determination module 603 is specifically used for: For any of the expert indicator sets, N candidate expert models are determined from the expert indicator set; the N candidate expert models are expert models whose performance indicators are ranked in the top N positions, N is an integer greater than 0, and the K target time series data samples correspond to K*N candidate expert models; Different candidate expert models among the K*N candidate expert models are determined as the target expert models.

[0098] Optionally, the target prediction module 604 includes: A candidate prediction unit, used for inputting the historical time series data into each target expert model to obtain candidate time series data of the target object in the target time period predicted by each target expert model; The weighted summation unit is used to perform weighted summation on the candidate time series data predicted by each target expert model based on the weight of each target expert model to obtain the target time series data.

[0099] Optionally, the above-mentioned device further comprises: A first determination module, configured to determine at least one performance indicator corresponding to each target expert model based on the performance indicators corresponding to each of the K*N candidate expert models; A second determination module, configured to determine a target performance indicator corresponding to each target expert model based on at least one performance indicator corresponding to each target expert model; The third determination module is used to determine the weight of each target expert model based on the target performance index corresponding to each target expert model.

[0100] Optionally, the above device further comprises: A prompt construction module, used for performing prompt construction on the historical time series data to obtain a first prompt word; The sample determination module 602 includes: a similarity determination unit, configured to determine a target similarity between each time series data sample and the historical time series data based on the historical time series data and the first prompt word; The target sample determination unit is used to determine the time series data samples ranked in the top K positions of the target similarity as the target time series data samples.

[0101] Optionally, the similarity determination unit is specifically used to: Based on the time series data samples and the historical time series data, calculating a first similarity between the time series data samples and the historical time series data; Calculating a second similarity between each time series data sample and the historical time series data based on the second prompt word of each time series data sample and the first prompt word; Based on the first similarity and the second similarity between each time series data sample and the historical time series data, a target similarity between each time series data sample and the historical time series data is determined.

[0102] Optionally, the above device further includes: A sample acquisition module, used to acquire the at least two time series data samples; A prediction value determination module, used for inputting each of the time series data samples into the M expert models respectively, to obtain the prediction value of the time series data sample output by each of the M expert models; A performance determination module, used to determine the performance indicators of the M expert models when performing time series prediction on the time series data samples based on the real values ​​and predicted values ​​of the time series data samples; An indicator set generation module is used to generate an expert indicator set corresponding to the time series data sample based on the M expert models and the performance indicators of the M expert models when performing time series prediction on the time series data sample.

[0103] Optionally, the above device further includes: The model acquisition module is used to obtain the target expert models from the cloud platform.

[0104] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0105] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 7 As shown, the electronic device 7 of this embodiment includes: at least one processor 70 ( Figure 7 Only one is shown in the figure), a memory 71, and a computer program 72 stored in the memory 71 and executable on the at least one processor 70, wherein the processor 70 implements the steps in the above method embodiment when executing the computer program 72.

[0106] The electronic device may include, but is not limited to, a processor 70 and a memory 71. Those skilled in the art will appreciate that Figure 7 It is only an example of the electronic device 7 and does not constitute a limitation on the electronic device 7. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0107] The processor 70 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0108] In some embodiments, the memory 71 may be an internal storage unit of the electronic device 7, such as a hard disk or memory of the electronic device 7. In other embodiments, the memory 71 may also be an external storage device of the electronic device 7, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 7. Further, the memory 71 may also include both an internal storage unit of the electronic device 7 and an external storage device. The memory 71 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory 71 may also be used to temporarily store data that has been output or is to be output.

[0109] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0110] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device that can carry the computer program code to the device / electronic device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electric carrier signal, a telecommunication signal, and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0111] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0112] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0113] In the embodiments provided in the present application, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0114] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0115] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A time series prediction method, characterized in that: include: Obtain historical time series data of the target object; Determine, from at least two time series data samples, K target time series data samples that match the historical time series data, where K is an integer greater than zero; Based on the expert indicator sets corresponding to the K target time series data samples, at least one target expert model matching the historical time series data is determined; the expert indicator set includes: a correspondence between M expert models and performance indicators, the performance indicators characterizing the performance of the corresponding expert model when performing time series prediction on the corresponding time series data samples, where M is an integer greater than 1; Based on the historical time series data and each target expert model, the target time series data of the target object in the target time period after the current moment is predicted.

2. The time series prediction method according to claim 1, characterized in that: The determining, based on the expert indicator sets corresponding to the K target time series data samples, at least one target expert model matching the historical time series data comprises: For any of the expert indicator sets, N candidate expert models are determined from the expert indicator set; the N candidate expert models are expert models whose performance indicators are ranked in the top N positions, N is an integer greater than 0, and the K target time series data samples correspond to K*N candidate expert models; Different candidate expert models among the K*N candidate expert models are determined as the target expert models.

3. The time series prediction method according to claim 2, characterized in that: The step of predicting target time series data of the target object in a target time period after the current moment based on the historical time series data and each target expert model includes: Inputting the historical time series data into each target expert model to obtain candidate time series data of the target object in the target time period predicted by each target expert model; Based on the weights of the target expert models, the candidate time series data predicted by the target expert models are weighted and summed to obtain the target time series data.

4. The time series prediction method according to claim 3, characterized in that: Before weighted summing of the candidate time series data predicted by the target expert models based on the weights of the target expert models, the method further includes: Determine at least one performance indicator corresponding to each target expert model based on the performance indicators corresponding to each of the K*N candidate expert models; Determining a target performance indicator corresponding to each target expert model based on at least one performance indicator corresponding to each target expert model; Based on the target performance indicators corresponding to the target expert models, the weights of the target expert models are determined.

5. The time series prediction method according to any one of claims 1 to 4, characterized in that: Before determining K target time series data samples matching the historical time series data from at least two time series data samples, the method further includes: Performing prompt construction on the historical time series data to obtain a first prompt word; The step of determining, from at least two time series data samples, K target time series data samples that match the historical time series data includes: Based on the historical time series data and the first prompt word, determining a target similarity between each time series data sample and the historical time series data; The time series data samples ranked in the top K positions of the target similarity are determined as the target time series data samples.

6. The time series prediction method according to claim 5, characterized in that: The determining, based on the historical time series data and the first prompt word, a target similarity between each time series data sample and the historical time series data includes: Based on the time series data samples and the historical time series data, calculating a first similarity between the time series data samples and the historical time series data; Calculating a second similarity between each time series data sample and the historical time series data based on the second prompt word of each time series data sample and the first prompt word; Based on the first similarity and the second similarity between each time series data sample and the historical time series data, a target similarity between each time series data sample and the historical time series data is determined.

7. The time series prediction method according to any one of claims 1 to 4, characterized in that: Before determining at least one target expert model matching the historical time series data based on the expert indicator sets corresponding to the K target time series data samples, the method further includes: Acquire the at least two time series data samples; For each of the time series data samples, input the time series data samples into the M expert models respectively to obtain the predicted values ​​of the time series data samples output by the M expert models respectively; Determine, based on the true value and the predicted value of the time series data sample, the performance indicators of the M expert models when performing time series prediction on the time series data sample; Based on the M expert models and the performance indicators of the M expert models when performing time series prediction on the time series data samples, an expert indicator set corresponding to the time series data samples is generated.

8. The time series prediction method according to any one of claims 1 to 4, characterized in that: Before predicting the target time series data of the target object in the target time period after the current moment based on the historical time series data and each target expert model, the method further includes: The target expert models are obtained from the cloud platform.

9. A time series prediction device, characterized in that: include: A data acquisition module is used to acquire historical time series data of a target object; A sample determination module, configured to determine K target time series data samples matching the historical time series data from at least two time series data samples, where K is an integer greater than zero; A target determination module, configured to determine at least one target expert model matching the historical time series data based on the expert indicator sets corresponding to the K target time series data samples; The expert indicator set includes: a correspondence between M expert models and performance indicators, wherein the performance indicators represent the performance of the corresponding expert model when performing time series prediction on the corresponding time series data sample, and M is an integer greater than 1; The target prediction module is used to predict the target time series data of the target object in a target time period after the current moment based on the historical time series data and each target expert model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the electronic device implements the time series prediction method according to any one of claims 1 to 8.

Images