Training method of time series prediction model, time series prediction method and related products

Abstract

Embodiments of the present application provide a training method of a time series prediction model, a time series prediction method and related products. The training method comprises: obtaining a sample time series at a specified time; inputting the sample time series into a time series feature extraction layer of a time series prediction model to be trained to extract a first time series; inputting the first time series into a time-frequency conversion layer of the time series prediction model to perform Fourier transform on the first time series and convert the first time series into a first frequency domain sequence with a specified sequence length, wherein the specified sequence length is smaller than a minimum sequence length of at least one first time series; inputting the first frequency domain sequence into an LSTM prediction layer of the time series prediction model to predict a second frequency domain sequence; inputting the second frequency domain sequence into a frequency-time conversion layer of the time series prediction model to perform inverse Fourier transform on the second frequency domain sequence to obtain a second time series; and training the time series prediction model based on a loss value of the second time series to obtain a trained time series prediction model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a training method for a time series prediction model, a time series prediction method, and related products. Background Technology

[0002] Time series forecasting is a method of predicting future trends based on the time series of historical statistical data. In related technologies, for large power-consuming institutions, such as large computer systems in large computing centers, time series forecasting can be used to predict trends in power consumption. Adjustments to the computer system can then be made based on these trends to fully utilize power resources and save electricity.

[0003] In existing technologies, LSTM (Long Short-Term Memory) models are commonly used for time series forecasting. However, for large computer systems in institutions such as computing centers, their power consumption exhibits long-term, seasonal, and periodic variations. This necessitates collecting long-term time series data for long-term power consumption forecasting. Long-term time series forecasting relies on large volumes of time series data, and existing LSTM models suffer from low forecasting efficiency and poor prediction results for long-term time series, making them unsuitable for power consumption forecasting in large computer systems such as computing centers. Summary of the Invention

[0004] This application provides a training method for a time series prediction model, a time series prediction method, and related products to improve the prediction efficiency of long-period time series.

[0005] In a first aspect, embodiments of this application provide a training method for a time series prediction model, the method comprising:

[0006] Obtain the sample time series at a specified time;

[0007] The sample time series is input into the time series feature extraction layer of the time series prediction model to be trained, and at least one first time series is extracted from the sample time series; wherein, the at least one first time series includes: the overall sample time series;

[0008] The at least one first time series is input into the time-frequency transformation layer of the time-series prediction model, and Fourier transform is performed on the at least one first time series to transform the first time series into a first frequency domain sequence of a specified sequence length, thereby obtaining at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0009] The at least one first frequency domain sequence is input into the LSTM prediction layer of the time series prediction model to predict at least one second frequency domain sequence of the target time series at the target time; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0010] The at least one second frequency domain sequence is input into the frequency-time transformation layer of the time series prediction model, and inverse Fourier transform is performed on the at least one second frequency domain sequence to obtain at least one second time series of the target time series; wherein, the at least one second time series corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time series includes: the overall target time series;

[0011] The time series prediction model is trained based on the loss value of the at least one second time series to obtain the trained time series prediction model.

[0012] Secondly, embodiments of this application provide a time series prediction method, the method comprising:

[0013] Get the specified time series at a specified time point to be predicted;

[0014] The specified time series is input into the trained time series prediction model to predict the target time series at the target time.

[0015] The time-series prediction model includes: a time-series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer;

[0016] The time-series feature extraction layer is used to extract at least one first time series from the specified time series; wherein, the at least one first time series includes: the overall time series of the specified time series;

[0017] The time-frequency transformation layer is used to perform Fourier transform on the at least one first time series, transforming the first time series into a first frequency domain sequence of a specified sequence length, to obtain at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0018] The LSTM prediction layer is used to predict at least one second frequency domain sequence of the target time series based on the at least one first frequency domain sequence; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0019] The frequency-time transformation layer is used to perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time sequence of the target time sequence; wherein, the at least one second time sequence corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time sequence includes: the overall target time sequence;

[0020] The time series prediction model is trained using the time series prediction model training method provided in the foregoing embodiments.

[0021] Thirdly, embodiments of this application provide a training apparatus for a time series prediction model, the apparatus comprising:

[0022] The sample acquisition module is used to acquire the sample time series at a specified time.

[0023] The feature extraction module is used to input the sample time series into the time series feature extraction layer of the time series prediction model to be trained, and extract at least one first time series from the sample time series; wherein, the at least one first time series includes: the overall sample time series;

[0024] The time-frequency transformation module is used to input the at least one first time series into the time-frequency transformation layer of the time-series prediction model, perform Fourier transform on the at least one first time series respectively, and transform the first time series into a first frequency domain sequence of a specified sequence length to obtain at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0025] The prediction module is used to input the at least one first frequency domain sequence into the LSTM prediction layer of the time series prediction model to predict at least one second frequency domain sequence of the target time series at the target time; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0026] The frequency-time transformation module is used to input the at least one second frequency domain sequence into the frequency-time transformation layer of the time series prediction model, and to perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time series of the target time series; wherein, the at least one second time series corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time series includes: the overall target time series of the target time series;

[0027] The training module is used to train the time series prediction model based on the loss value of the at least one second time series to obtain the trained time series prediction model.

[0028] Fourthly, embodiments of this application provide a timing prediction apparatus, the apparatus comprising:

[0029] The sequence acquisition module is used to acquire a specified time series at a specified time point to be predicted;

[0030] The prediction module is used to input the specified time series into the trained time series prediction model to predict the target time series at the target time.

[0031] The time-series prediction model includes: a time-series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer;

[0032] The time-series feature extraction layer is used to extract at least one first time series from the specified time series; wherein, the at least one first time series includes: the overall time series of the specified time series;

[0033] The time-frequency transformation layer is used to perform Fourier transform on the at least one first time series respectively, transforming the first time series into a first frequency domain sequence of a specified sequence length, to obtain at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0034] The LSTM prediction layer is used to predict at least one second frequency domain sequence of the target time series based on the at least one first frequency domain sequence; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0035] The frequency-time transformation layer is used to perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time sequence of the target time sequence; wherein, the at least one second time sequence corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time sequence includes: the overall target time sequence of the target time sequence;

[0036] The time series prediction model is trained using the time series prediction model training method provided in the foregoing embodiments.

[0037] Fifthly, embodiments of this application provide a computer device, the device comprising:

[0038] Processor; and

[0039] A memory is configured to store computer-executable instructions configured to be executed by the processor, the executable instructions including steps for performing a training method for a timing prediction model as provided in the foregoing embodiments or a timing prediction method as provided in the foregoing embodiments.

[0040] Sixthly, embodiments of this application provide a storage medium for storing computer-executable instructions that cause a computer to execute a training method for a timing prediction model as provided in the foregoing embodiments or a timing prediction method as provided in the foregoing embodiments. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in one or more of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart illustrating a training method for a time series prediction model provided in an embodiment of this application;

[0043] Figure 2 for Figure 1 A schematic diagram of the structure of the time series prediction model to be trained in the training method shown;

[0044] Figure 3 A flowchart illustrating a time series prediction method provided in an embodiment of this application;

[0045] Figure 4 A schematic diagram of a training device for a time series prediction model provided in an embodiment of this application;

[0046] Figure 5 A schematic diagram of a timing prediction device provided in an embodiment of this application;

[0047] Figure 6 This is a schematic diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0048] To enable those skilled in the art to better understand the technical solutions in one or more of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of one or more of this application, and not all embodiments. Based on the embodiments of one or more of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this application.

[0049] It should be noted that, unless otherwise specified, one or more embodiments and features described in this application can be combined with each other. The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0050] As described in the background art, LSTM has low prediction efficiency and poor prediction performance for predicting long-period time series such as the power consumption of large computing centers. This application provides a training scheme and a time series prediction scheme for a time series prediction model. This scheme extracts the overall time series from a long-period time series at a specified time, performs a Fourier transform on the overall time series, compresses the long-period time series into a short-period frequency domain sequence of a specified length, predicts the short-period frequency domain sequence of the target time series based on the short-period frequency domain sequence corresponding to the long-period time series at the specified time, and then performs an inverse Fourier transform on the short-period frequency domain sequence of the target time series to obtain the target time series. Because the long-period time series is compressed in the frequency domain, the amount of data processed for prediction can be reduced while preserving the sequence variation characteristics of the long-period time series, thus improving prediction efficiency. Furthermore, this application embodiment compresses the overall long-period time series into... While setting a frequency domain sequence of a certain length, one or more local time series with significant or recent changes are obtained from the long-period time series. Fourier transforms are then performed on these local time series, compressing them into short-period frequency domain sequences. The short-period frequency domain sequences corresponding to both the overall and local time series of the long-period time series are combined as multi-dimensional frequency domain features of the long-period time series. Prediction of the long-period time series is then performed based on these multi-dimensional frequency domain features. This approach captures both the overall and local features of the long-period time series while reducing the amount of data processed for prediction. The model adapts to the long and short-period characteristics of the long-period time series, improving both the efficiency and accuracy of long-period time series prediction.

[0051] Figure 1 This application provides a training method for a time series prediction model. Figure 2 for Figure 1 The diagram shows the structure of the time-series prediction model described in the training method. (Refer to...) Figure 1 , Figure 2 The training method of the time series prediction model provided in the embodiments of this application will be described in detail.

[0052] The training method for the time series prediction model provided in this application embodiment is used to train a time series prediction model, which is used for predicting long-period time series. Its structure is as follows: Figure 2 As shown. Figure 2As shown, the time series prediction model includes: a time series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer. The time series feature extraction layer extracts at least one first time series from a specified time series at a specified input time. The time-frequency transformation layer performs a Fourier transform on the at least one first time series, transforming it into a first frequency domain sequence of a specified length, resulting in at least one first frequency domain sequence, where the specified sequence length is less than the minimum sequence length of the at least one first time series. The LSTM prediction layer, including an LSTM network, predicts at least one second frequency domain sequence of the target time series based on the at least one first frequency domain sequence. The frequency-time transformation layer performs an inverse Fourier transform on the at least one second frequency domain sequence, obtaining at least one second time series of the target time series. Both the specified time series and the target time series are long-period time series, and their sequence lengths can be the same. The training process of this time series prediction model is as follows: Figure 1 As shown, Figure 1 As shown, the training method for the time series prediction model provided in this application may include the following steps:

[0053] S102, obtain the sample time series at the specified time.

[0054] A time series is a sequence of values ​​for the same statistical indicator arranged in chronological order of their occurrence. It is typically the result of observing the values ​​of the indicator at equal intervals according to a given sampling rate. A sample time series at a specified time is a long-period time series corresponding to that specified time. For example, a time series with a specified time endpoint (i.e., the end point of the time series) is a time series with a specified time interval and a specified time length.

[0055] In applications where power consumption of a large computing center's mainframe computer system is adjusted based on its power consumption to save energy, adjustments to the system's computing power and capacity are typically involved. The computing power and capacity of a computer system are strongly correlated with the quality of service it provides, thus requiring more precise adjustment strategies. To achieve this, long-term, fine-grained power consumption forecasting is necessary. Therefore, when forecasting the power consumption of a large computing center's mainframe computer system, long-term, fine-grained forecasting is required. In this embodiment, taking the application scenario of forecasting the power consumption of a large computing center's mainframe computer system as an example, the sample time series can be a fine-grained time series of a set length obtained by sampling the mainframe computer system's power consumption at a set sampling interval. If the time interval of the time series is less than a set time interval threshold, the time series is considered a fine-grained time series. For example, the sample time series can be a power consumption series with a specified endpoint, a time interval of one minute, and a time length of one year; that is, the power consumption value per minute for one year prior to the specified endpoint.

[0056] In implementation, a large computer system in a large computing center can be used as the target computer system. Power consumption of the target computer system is collected at set sampling intervals to obtain a time series of its power consumption. A sample time series of power consumption ending at a specified time is obtained from the time series of power consumption of the target computer system through a time window of set length. This specified time can be any time, such as time T1. The set time length can be determined based on the periodicity of the regular changes in the power consumption of the target computer system, such as one year. The set sampling interval can be determined based on the desired granularity of predicting the power consumption of the target computer system, such as one minute.

[0057] S104, input the sample time series into the time series prediction model to be trained, and extract at least one first time series from the sample time series.

[0058] The time-series feature extraction layer is used to extract at least one first time series from the sample time series. The at least one first time series includes at least the overall sequence of the sample time series, i.e., the overall sample time series.

[0059] In one implementation, at least one first time series includes at least two first time series, one of which can be the overall sequence of the sample time series, i.e., the overall sample time series, and the other time series can be a local sequence of the sample time series, i.e., the local sample time series. This allows for the capture of long-term and short-term variations in long-period sample time series. By training the time series prediction model based on these long-term and short-term variations, the model's ability to capture long-term and short-term information in long-period time series is improved, thereby enhancing the prediction accuracy of the time series prediction model for long-period time series.

[0060] In implementation, the time-series feature extraction layer can further include a global feature extraction network and a local feature extraction network. The global feature extraction network extracts the overall sample time series from the sample time series, i.e., it extracts the overall temporal features of the sample time series. The local feature extraction layer extracts the local sample time series from the sample time series, i.e., it extracts the local temporal features of the sample time series. After the sample time series is input into the time-series feature extraction layer, the global feature extraction layer can extract the overall sample time series, and the local feature extraction layer can extract the local sample time series.

[0061] As an optional implementation, at least one first time series may include a global sample time series and one or more local sample time series. The global sample time series may be the entire time series of the sample time series, and the local sample time series may be the parts of the sample time series that have significant changes. In this way, the overall long-term information and the local information of significant changes of the sample time series can be obtained, enabling the time series prediction model to adaptively learn the long-term change information and local significant change information of the sample time series, thereby improving the model's ability to capture the long-term change information and significant change information of long-term time series, and thus improving the prediction ability of long-term time series.

[0062] In implementation, the local feature extraction network can adopt a salient feature extraction network, which is used to extract the parts of the time series that have undergone significant changes, such as local time series with a set length and a variance value exceeding a set threshold. Thus, the salient feature extraction network can extract the significant time series that have undergone significant changes from the sample time series, thereby obtaining one or more local sample time series.

[0063] In one implementation, the salient feature extraction network may include a sliding window and a decision unit. The sliding window is used to select candidate local time series, which can be local time series within the sample time series with a sequence length of a first predetermined length, where the first predetermined length is less than the total sequence length of the sample time series. The decision unit is used to calculate the standard deviation or variance of the candidate local time series, and when the standard deviation or variance of the local time series exceeds a predetermined threshold, it determines the candidate local time series as the target local time series and extracts the target local time series as the first time series.

[0064] Taking the power consumption of a large computer server system as an example, the sample time series can be a time series of power consumption with a specified time endpoint, a time interval of one minute, and a time length of one year, that is, the power consumption of one minute in the year before the specified time. At least one first time series obtained from the sample time series can include: a global sample time series and one or more local sample time series. The global sample time series can be a complete sample time series, that is, a time series of power consumption with a specified time endpoint, a time interval of one minute, and a time length of one year. The local sample time series can be a time series of a certain month in which the variance of power consumption in the sample time series exceeds a set threshold.

[0065] As an optional implementation, at least one first time series may include a global sample time series and one or more local sample time series. The global sample time series can be the entire time series of the sample time series, and the local sample time series can be the most recent time series within the sample time series. The most recent time series is a portion of the sample time series whose endpoint is the same as the global sample time series, but whose sequence length is less than the length of the global sample time series. This allows for the acquisition of both long-term overall change information and short-term recent change information of the sample time series, enabling the time series prediction model to adaptively learn from these two information. This improves the model's ability to capture both long-term and short-term changes in long-period time series, thereby enhancing its predictive power for long-period time series.

[0066] In implementation, the local feature extraction layer can employ a recent feature extraction network. This network extracts local time series with a specified endpoint and a set time length from the overall time series. This set time length is shorter than the overall time series. Thus, one or more recent time series can be extracted from the sample time series, resulting in one or more local sample time series. Multiple recent time series within the sample time series share the same endpoint as the overall time series, but their sequence lengths are shorter than the overall time series. Furthermore, any two recent time series have different sequence lengths, allowing for the acquisition of recent change information of different lengths within the sample time series.

[0067] Taking the power consumption of a large computer server system as an example, the sample time series can be a time series of power consumption with a specified end time, a time interval of one minute, and a time length of one year, that is, the power consumption of one minute one year before the specified end time. At least one first time series obtained from the sample time series can include: a global sample time series and one or more local sample time series. The global sample time series can be a complete sample time series, that is, a time series of power consumption with a specified end time, a time interval of one minute, and a time length of one year. The local sample time series can be a recent time series of the sample time series. Multiple local sample time series can include: the power consumption per minute one month before the specified end time, that is, the power consumption per minute three months before the specified end time, that is, the power consumption per minute six months before the specified end time, that is, the power consumption per minute six months before the specified end time.

[0068] S106, input at least one first time series into the time-frequency transformation layer of the time-series prediction model, perform Fourier transform on the at least one first time series respectively, and transform the first time series into a first frequency domain sequence of a specified sequence length.

[0069] The time-frequency transformation layer is used to transform the first time series into a first frequency domain sequence of a specified sequence length, wherein the specified sequence length is less than the minimum sequence length in at least one of the first time series.

[0070] For long-period time series, the data volume is enormous, resulting in a large amount of data processing and poor prediction performance. In this embodiment, a Fourier transform is used to transform the long-period first time series into a short-period first frequency domain series. This time-frequency conversion preserves the variation information of the long-period time series while compressing the data volume, thereby reducing the amount of data processed for prediction and improving prediction efficiency.

[0071] It is understandable that regardless of the length or volume of the time series, it can be converted into a frequency domain sequence of a specified length using Fourier transform, thereby achieving data compression. In one implementation, Fourier transforms can be performed on at least one first time series with different lengths, transforming them all into first frequency domain sequences of a specified length. These sequences are then used as frequency domain representations of the sample time series for sequence prediction. The specified sequence length is less than the shortest among the lengths of the multiple first time series. This allows the model to capture the long and short period variations of long-period time signals while simultaneously achieving feature compression, thus improving the model's prediction accuracy.

[0072] The first frequency domain sequence is derived from the first time series through Fourier transform. It is the frequency domain sequence corresponding to the first time series. By performing Fourier transforms on multiple first time series respectively, the corresponding first frequency domain sequences can be obtained, resulting in multiple first frequency domain sequences, each corresponding to a different first time series. Furthermore, since the first time series is transformed into a first frequency domain sequence of a specified length during the time-frequency transformation, the sequence length of each of the multiple first frequency domain sequences is the specified length, and any two first frequency domain sequences have the same sequence length.

[0073] In implementation, the time-frequency conversion between the first time series and the first frequency domain sequence can be performed using a Fast Fourier Transform (FFT) with a set number of points. Specifically, a Fast Fourier Transform with a preset number of points can be performed on multiple first time series, transforming each of the multiple first time series into a first frequency domain sequence with a preset number of points, resulting in multiple first frequency domain sequences. Each of the multiple first frequency domain sequences corresponds one-to-one with the multiple first time series, and any two of the multiple first frequency domain sequences have the same sequence length, which is the preset number of points, less than the sequence length of any single first time series.

[0074] S108, at least one first frequency domain sequence is input into the LSTM prediction layer of the time series prediction model to obtain multiple second frequency domain sequences.

[0075] The LSTM prediction layer is used to predict at least one second frequency domain sequence of the target time series at a specified time based on at least one first frequency domain sequence of the sample time series. Each second frequency domain sequence corresponds one-to-one with at least one first frequency domain sequence. The target time series and the sample time series can have the same sequence length, and both can be long-period time series. Multiple second frequency domain sequences represent the frequency domain of multiple second time series of the target time series at the target time, and each second frequency domain sequence can be converted into its corresponding second time series after an inverse Fourier transform.

[0076] Taking the prediction of power consumption of a large computer server system as an example, the first frequency domain sequence of a sample time series with a specified endpoint (e.g., time T1), a duration of one year, and a time interval of one minute can be input into the LSTM prediction layer. The second frequency domain sequence of the target time series with a target endpoint (e.g., time T1+T), a duration of one year, and a time interval of one minute can be predicted. Here, the second frequency domain sequence is the frequency domain representation of the second time series of the target time series, and the second time series can be either the overall target time series or a partial target time series.

[0077] In one implementation, since at least one first frequency domain sequence has the same sequence length, each first frequency domain sequence is a frequency domain representation of its corresponding first time series, and each first time series represents the overall change characteristics or local change characteristics of the sample time series. Therefore, each first frequency domain sequence can represent the overall change characteristics or local change characteristics of the sample time series. Each first frequency domain sequence can be used as a frequency domain feature of one dimension. Multiple first frequency domain sequences can be merged into a first multi-dimensional frequency domain feature matrix. This first multi-dimensional frequency domain feature matrix can characterize the overall frequency domain features and one or more local frequency domain features of the sample time series at a specified time. Then, the first multi-dimensional frequency domain feature matrix is ​​input into the LSTM prediction layer of the time series prediction model to predict the second multi-dimensional frequency domain feature matrix that characterizes the frequency domain features of the target time series at the target time. The second multi-dimensional frequency domain feature matrix is ​​a merged matrix of multiple second frequency domain sequences of the target time series at the target time. Each second frequency domain sequence can be converted into its corresponding second time series after inverse Fourier transform. The second time series is the overall target time series or a local target time series of the target time series. The second multi-dimensional frequency domain feature matrix can characterize the overall frequency domain features and one or more local frequency domain features of the target time series at the target time.

[0078] S110, input at least one second frequency domain sequence into the frequency-time transformation layer of the time series prediction model, and perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time series.

[0079] The frequency-time transformation layer performs an inverse Fourier transform on at least one second frequency domain sequence to obtain at least one second time series of the target time series. This at least one second time series includes: the overall target time series of the target time series, which is obtained by processing the overall sample time series of the sample time series through a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer. Both the overall target time series and the overall sample time series can be long-period time series, and their sequence lengths can be the same, for example, both one year. Their sequence endpoints, i.e., the time endpoints of the sequences, are different. The time endpoint of the overall target time series can be a target time, while the time endpoint of the overall sample time series can be a specified time.

[0080] It is understood that at least one second time series corresponds one-to-one with at least one second frequency domain sequence. Each second time series is derived from its corresponding second frequency domain sequence through inverse Fourier transform. The second time series and the second frequency domain sequence correspond one-to-one, and the second frequency domain sequence corresponds one-to-one with the first frequency domain sequence, and the first frequency domain sequence corresponds one-to-one with the first time series. That is, the first time series, the first frequency domain sequence, the second frequency domain sequence, and the second time series are all in a one-to-one correspondence. Each second time series is obtained by processing its corresponding first time series through a time-frequency transform layer, an LSTM prediction layer, and a frequency-time transform layer. Therefore, when the first time series includes a local sample time series, the second time series includes the local target time series corresponding to that local sample time series. The two have different time endpoints, but the sequence lengths can be the same.

[0081] Taking the prediction of power consumption of a large computer server system as an example, if the first time series is a local sample time series with a time end point of T1, a time interval of one minute, and a time length of 15 days, the corresponding second time series can be a local target time series with a time end point of T1+T and a time length of 15 days; if the first time series is a local sample time series with a time end point of T1, a time interval of one minute, and a time length of one month, the corresponding second time series can be a local target time series with a time end point of T1+T, a time interval of one minute, and a time length of one month.

[0082] S112, Based on the loss value of at least one second time series, the time series prediction model is trained to obtain the trained time series prediction model.

[0083] The sample time series can be a labeled sequence. The loss value of the second time series can be calculated based on the labels of the second time series and the sample time series. Then, the time series prediction model is trained in a supervised manner based on the loss value of the second time series. Of course, it is understandable that the sample time series can also be an unlabeled sequence, and the time series prediction model can still be trained in an unsupervised manner.

[0084] In one implementation, there can be only one of the first time series and the second time series. The first time series can be the overall sample time series, and the second time series can be the overall target time series. The loss value of the overall target time series can be calculated based on the label values ​​of the overall target time series and the sample time series. The model is trained based on the loss value of the overall target time series.

[0085] In one implementation, there can be multiple first and second time series. The first time series may include an overall sample time series and one or more local sample time series. Correspondingly, the second time series may include an overall target time series and one or more local target time series. The loss value of the target time series can be calculated based on the loss values ​​of the multiple second time series using a preset loss function. Based on the loss value of the target time series, the parameters of the time series prediction model are updated using a backpropagation algorithm, thereby training the model. The preset loss function can be a weighted sum of the loss functions corresponding to at least one second time series. Similarly, the loss value of the target time series can be a weighted sum of the loss values ​​corresponding to at least one second time series. The weights of the loss functions corresponding to each second time series can be adjusted as hyperparameters during training. The loss functions corresponding to the second time series can be existing loss functions, such as the cross-entropy loss function.

[0086] This application embodiment extracts at least one first time series from a long-period sample time series. The at least one first time series includes: an overall sample time series; performing a Fourier transform on the overall sample time series to compress the long-period overall sample time series into a first frequency domain sequence of a specified sequence length, where the specified sequence length is less than the minimum sequence length of at least one first time series; and performing prediction and training based on the first frequency domain sequence. This can reduce the amount of data processed for prediction while preserving the overall sequence variation characteristics of the long-period time series, thereby improving prediction efficiency. Furthermore, this application embodiment also includes at least one first time series extracted from the long-period sample time series: a local time series, which is also compressed into a first frequency domain sequence of a specified sequence length. A frequency domain sequence is used, where the first frequency domain sequences corresponding to the overall sample time series and the local sample time series are used as the long and short period frequency domain features of the sample time series. Based on the long and short period frequency domain features of the sample time series, the long and short period frequency domain features of the target time series are predicted. The long and short period time domain features of the target time series are obtained through the long and short period frequency domain features of the target time series. The model is trained based on the common loss of the long and short period time domain features of the target time series. This can reduce the amount of data to be processed for prediction while enabling the model to capture the long and short period variation information of the long period time series. The model can adapt to the long and short period features of the long period time series, thereby improving the prediction efficiency and accuracy of the long period time series.

[0087] Furthermore, embodiments of this application also provide a time series prediction method. Figure 3 The following is a flowchart illustrating a time series prediction method provided in an embodiment of this application, as shown below. Figure 3 As shown, the timing prediction method provided in this application embodiment may include the following steps:

[0088] S202, obtain the specified time series at the specified time point to be predicted.

[0089] The specified time series can be a long-period time series, and its time endpoint can be a specified moment.

[0090] Taking the prediction of power consumption of large computer systems in a large computing center as an example, the specified time series can be a power consumption series with a specified time endpoint, a time interval of one minute, and a time length of one year, that is, a series of power consumption values ​​per minute for the year prior to the specified time.

[0091] S204, Input the specified time series into the trained time series prediction model to predict the target time series at the target time.

[0092] The target time series can be a long-period time series, and its time endpoint can be the target time.

[0093] Taking the prediction of power consumption of large computer systems in large computing centers as an example, the target time series can be a power consumption series with the target time as the end point, a time interval of one minute, and a time length of one year, that is, a series of power consumption values ​​per minute in the year before the target time.

[0094] The structure of the time series prediction model is as follows: Figure 2 As shown, it includes: a time-series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer; wherein, the time-series feature extraction layer is used to extract at least one first time series from the specified time series; wherein, the at least one first time series includes: the overall time series of the specified time series; the time-frequency transformation layer is used to perform a Fourier transform on the at least one first time series, transforming the first time series into a first frequency domain sequence of a specified sequence length, to obtain at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence... The columns correspond one-to-one with the at least one first time series; the LSTM prediction layer is used to predict at least one second frequency domain sequence of the target time series based on the at least one first frequency domain sequence; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence; the frequency-time transformation layer is used to perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time series of the target time series; wherein, the at least one second time series corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time series includes: the overall target time series of the target time series.

[0095] In one implementation, the at least one first time series further includes: a local time series of the specified time series; extracting at least one first time series from the specified time series includes: extracting the overall time series of the specified time series from the specified time series; and extracting one or more local time series of the specified time series from the specified time series.

[0096] In one implementation, the local time series is a significant change time series or a recent time series of the specified time series.

[0097] In one implementation, the specified time series is a time series of the power consumption of the target computer system at a specified time, and the target time series is a time series of the power consumption of the target computer system at a target time; the method further includes: determining the power consumption change of the target computer system in a target time period based on the target time series; and adjusting the power allocation of the target computer system based on the power consumption change of the target computer system in the target time period.

[0098] The target computer system can be a cloud computing system, and its power allocation can be adjusted by adjusting its computing capacity. Therefore, in one implementation, adjusting the power allocation of the target computer system based on changes in its power consumption during the target time period includes:

[0099] When the power consumption of the target computer system during the target time period shows an increasing trend, the capacity of the target computer system is expanded to reduce the power consumption of the target computer system during the target time period.

[0100] When the power consumption of the target computer system during the target time period shows a decreasing trend, the target computer system is downsized to increase its power consumption during the target time period.

[0101] By adjusting the power allocation of the target computer system according to the changing trend of its power consumption during the target time period, the target computer system can make full use of the power during the target time period, thereby improving the efficiency of power resource utilization and reducing unnecessary power loss.

[0102] The time series prediction method provided in this application extracts the overall time series from a long-period time series at a specified time. It then performs a Fourier transform on the overall time series of the long-period time series, compressing it into a short-period frequency domain sequence of a specified length. Based on the short-period frequency domain sequence corresponding to the long-period time series at the specified time, it predicts the short-period frequency domain sequence of the target time series. Finally, it performs an inverse Fourier transform on the short-period frequency domain sequence of the target time series to obtain the target time series. Because the long-period time series is compressed in the frequency domain, it can reduce the amount of data processed for prediction while preserving the sequence variation characteristics of the long-period time series, thus improving prediction efficiency. Furthermore, this application embodiment, in the process of compressing the overall time series of the long-period time series... While compressing the data into a frequency domain sequence of a set length, one or more local time series with significant or recent changes are extracted from the long-period time series. Fourier transforms are then performed on these local time series, which are also compressed into short-period frequency domain sequences. The short-period frequency domain sequences corresponding to both the overall and local time series of the long-period time series are used together as multi-dimensional frequency domain features of the long-period time series. Predictions of long-period target time series are then made based on these multi-dimensional frequency domain features. This approach captures both the overall and local features of the long-period time series while reducing the amount of data processed for prediction, thus improving both the efficiency and accuracy of long-period time series prediction.

[0103] It is understood that the trained time series prediction model used in the embodiments of this application can be trained using the training method of the time series prediction model provided in the above embodiments. The specific training method can be referred to the foregoing embodiments, and will not be repeated here.

[0104] Corresponding to the training method of the time series prediction model provided in the above embodiments, based on the same technical concept, this application embodiment also provides a training device for the time series prediction model, which is used to execute the training method of the time series prediction model provided in the above embodiments. Figure 4 This is a schematic diagram of the modules of the training device for the time series prediction model provided in the embodiments of this application, as shown below. Figure 4 As shown, the training device includes: a sample acquisition module 402, a feature extraction module 404, a time-frequency transformation module 406, a prediction module 408, a time-frequency transformation module 410, and a training module 412, wherein:

[0105] The sample acquisition module 402 is used to acquire the sample time series at a specified time.

[0106] The feature extraction module 404 is used to input the sample time series into the time series feature extraction layer of the time series prediction model to be trained, and extract at least one first time series from the sample time series; wherein, the at least one first time series includes: the overall sample time series;

[0107] The time-frequency transformation module 406 is used to input the at least one first time series into the time-frequency transformation layer of the time series prediction model, perform Fourier transform on the at least one first time series respectively, and transform the first time series into a first frequency domain sequence of a specified sequence length to obtain at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0108] Prediction module 408 is used to input the at least one first frequency domain sequence into the LSTM prediction layer of the time series prediction model to predict at least one second frequency domain sequence of the target time series at the target time; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0109] The frequency-time transformation module 410 is used to input the at least one second frequency domain sequence into the frequency-time transformation layer of the time series prediction model, and perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time series of the target time series; wherein, the at least one second time series corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time series includes: the overall target time series of the target time series.

[0110] Training module 412 is used to train the time series prediction model based on the loss value of the at least one second time series to obtain the trained time series prediction model.

[0111] In one implementation, the at least one first time series further includes: a local sample time series of the sample time series;

[0112] The step of extracting at least one first time series from the sample time series includes:

[0113] Extract the overall sample time series from the sample time series;

[0114] One or more local sample time series are extracted from the sample time series using a salient feature extraction network; wherein the local sample time series are the parts of the sample time series that have undergone significant changes.

[0115] In one implementation, the at least one first time series further includes: a local time series of the sample time series; extracting at least one first time series from the sample time series includes:

[0116] Extract the overall sample time series from the sample time series;

[0117] One or more local sample time series are extracted from the sample time series using a recent feature extraction network; wherein, the local sample time series is a recent time series in the sample time series that has the same time end point as the overall sample time series.

[0118] In one implementation, performing Fourier transforms on the at least one first time series to transform the first time series into a first frequency domain sequence of a specified length, thereby obtaining at least one first frequency domain sequence, includes:

[0119] A preset number of Fast Fourier Transforms are performed on the multiple first time series to transform them into a preset number of first frequency domain sequences, resulting in multiple first frequency domain sequences; wherein, the multiple first frequency domain sequences correspond one-to-one with the multiple first time series, and any two first frequency domain sequences in the multiple first frequency domain sequences have the same sequence length.

[0120] In one implementation, the step of inputting the at least one first frequency domain sequence into the LSTM prediction layer of the time series prediction model to predict at least one second frequency domain sequence of the target time series at the target time includes:

[0121] Using each first frequency domain sequence as a feature of one dimension, the multiple first frequency domain sequences are merged into a first multi-dimensional frequency domain feature matrix; wherein, the first multi-dimensional frequency domain feature matrix represents the multi-dimensional frequency domain features of the sample time series at the specified time.

[0122] The first multi-dimensional frequency domain feature matrix is ​​input into the LSTM prediction layer of the time series prediction model to predict the second multi-dimensional frequency domain feature matrix; wherein, the second multi-dimensional frequency domain feature matrix represents the multi-dimensional frequency domain features of the target time series at the target time, and it is a merged matrix of multiple second frequency domain sequences of the target time series at the target time.

[0123] In one implementation, the step of inputting the at least one second frequency domain sequence into the frequency-time transform layer of the time series prediction model, and performing inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time series of the target time series includes:

[0124] The second multi-dimensional frequency domain feature matrix is ​​input into the frequency-time transformation layer of the time series prediction model, and the multiple second frequency domain sequences in the second multi-dimensional frequency domain feature matrix are subjected to inverse Fourier transform to obtain multiple second time series; wherein, the multiple second time series include: the overall target time series and the local target time series of the target time series.

[0125] In one implementation, training the time series prediction model based on the loss value of the at least one second time series to obtain the trained time series prediction model includes:

[0126] Based on a preset loss function, the loss value of the target time series is calculated according to the loss values ​​of the at least one second time series;

[0127] Based on the loss value of the target time series, the parameters of the time series prediction model are updated using the backpropagation algorithm.

[0128] In one implementation, the sample time series is a time series of electricity consumption.

[0129] The step of obtaining the sample time series at a specified time includes:

[0130] The power consumption of the target computer system is collected according to the set sampling interval to obtain the time series of the power consumption of the target computer system;

[0131] By setting a time window of a certain length, a sample time series of power consumption with a specified time endpoint is obtained from the time series of power consumption of the target computer system.

[0132] It should be noted that the training device for the time series prediction model provided in this application embodiment and the training method for the time series prediction model provided in this application embodiment are based on the same inventive concept. Therefore, for the specific implementation of the training device for the time series prediction model embodiment, please refer to the implementation of the corresponding training method for the time series prediction model mentioned above. Repeated parts will not be described again.

[0133] Corresponding to the timing prediction method provided in the above embodiments, based on the same technical concept, this application also provides a timing prediction device, which is used to execute the timing prediction method provided in the above embodiments. Figure 5 This is a schematic diagram of the timing prediction device provided in the embodiments of this application, as shown below. Figure 5 As shown, the time series prediction device includes a sequence acquisition module 502 and a prediction module 504, wherein:

[0134] The sequence acquisition module 502 is used to acquire a specified time series at a specified time to be predicted;

[0135] Prediction module 504 is used to input the specified time series into the trained time series prediction model to predict the target time series at the target time.

[0136] The time-series prediction model includes: a time-series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer;

[0137] The time-series feature extraction layer is used to extract at least one first time series from the time series at the specified time; wherein, the at least one first time series includes: the overall time series of the time series at the specified time;

[0138] The time-frequency transformation layer is used to perform Fourier transform on the at least one first time series respectively, transforming the first time series into a first frequency domain sequence of a specified sequence length, to obtain at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0139] The LSTM prediction layer is used to predict at least one second frequency domain sequence of the target time series based on the at least one first frequency domain sequence; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0140] The frequency-time transformation layer is used to perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time sequence of the target time sequence; wherein, the at least one second time sequence corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time sequence includes: the overall target time sequence of the target time sequence;

[0141] The time series prediction model is trained using the time series prediction model training method provided in the foregoing embodiments.

[0142] In one implementation, the specified time series is a time series of the power consumption of the target computer system at a specified time, and the target time series is a time series of the power consumption of the target computer system at a target time.

[0143] The device further includes:

[0144] The electricity consumption trend determination module is used to determine the change in electricity consumption of the target computer system during the target time period based on the target time series.

[0145] A power adjustment module is used to adjust the power distribution of the target computer system according to the changes in power consumption of the target computer system during the target time period.

[0146] In one implementation, the power adjustment module includes:

[0147] An expansion unit is used to expand the capacity of the target computer system when the trend of the power consumption of the target computer system during the target time period is increasing, so as to reduce the power consumption of the target computer system during the target time period.

[0148] A capacity reduction unit is used to reduce the capacity of the target computer system when the trend of power consumption of the target computer system during a target time period is decreasing, so as to make full use of the power of the target computer system during the target time period.

[0149] It should be noted that the timing prediction device and the timing prediction method provided in this application are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the corresponding timing prediction method described above, and the repeated parts will not be described again.

[0150] Furthermore, corresponding to the training method and time series prediction method of the time series prediction model provided in the above embodiments, based on the same technical concept, this application also provides a computer device, which is used to execute the above-described training method of the time series prediction model or to execute the above-described time series prediction method, such as... Figure 6 As shown.

[0151] Computer devices can vary significantly in configuration and performance, and may include one or more processors and memory. The memory may store one or more application programs or data. The memory can be temporary or permanent storage. The application programs stored in the memory may include one or more modules, each of which may include a series of computer-executable instructions for the computer device. Furthermore, the processor may be configured to communicate with the memory and execute the series of computer-executable instructions stored in the memory on the computer device. Computer devices may also include one or more power supplies, one or more wired or wireless network interfaces, one or more input / output interfaces, one or more keyboards, etc.

[0152] In one specific embodiment, the computer device includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for use in the computer device, and is configured to be executed by one or more processors. The one or more programs include computer-executable instructions for performing the following:

[0153] Obtain the sample time series at a specified time;

[0154] The sample time series is input into the time series feature extraction layer of the time series prediction model to be trained, and at least one first time series is extracted from the sample time series; wherein, the at least one first time series includes: the overall sample time series;

[0155] The at least one first time series is input into the time-frequency transformation layer of the time-series prediction model, and Fourier transform is performed on the at least one first time series to transform the first time series into a first frequency domain sequence of a specified sequence length, thereby obtaining at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0156] The at least one first frequency domain sequence is input into the LSTM prediction layer of the time series prediction model to predict at least one second frequency domain sequence of the target time series at the target time; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0157] The at least one second frequency domain sequence is input into the frequency-time transformation layer of the time series prediction model, and inverse Fourier transform is performed on the at least one second frequency domain sequence to obtain at least one second time series of the target time series; wherein, the at least one second time series corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time series includes: the overall target time series of the target time series;

[0158] The time series prediction model is trained based on the loss value of the at least one second time series to obtain the trained time series prediction model.

[0159] In one implementation, the at least one first time series further includes: a local sample time series of the sample time series;

[0160] The step of extracting at least one first time series from the sample time series includes:

[0161] Extract the overall sample time series from the sample time series;

[0162] One or more local sample time series are extracted from the sample time series using a salient feature extraction network; wherein the local sample time series are the parts of the sample time series that have undergone significant changes.

[0163] In one implementation, the at least one first time series further includes: a local time series of the sample time series;

[0164] The step of extracting at least one first time series from the sample time series includes:

[0165] Extract the overall sample time series from the sample time series;

[0166] One or more local sample time series are extracted from the sample time series using a recent feature extraction network; wherein, the local sample time series is a recent time series in the sample time series that has the same time end point as the overall sample time series.

[0167] In one implementation, performing Fourier transforms on the at least one first time series to transform the first time series into a first frequency domain sequence of a specified length, thereby obtaining at least one first frequency domain sequence, includes:

[0168] A preset number of Fast Fourier Transforms are performed on the multiple first time series to transform them into a preset number of first frequency domain sequences, resulting in multiple first frequency domain sequences; wherein, the multiple first frequency domain sequences correspond one-to-one with the multiple first time series, and any two first frequency domain sequences in the multiple first frequency domain sequences have the same sequence length.

[0169] In one implementation, the step of inputting the at least one first frequency domain sequence into the LSTM prediction layer of the time series prediction model to predict at least one second frequency domain sequence of the target time series at the target time includes:

[0170] Using each first frequency domain sequence as a feature of one dimension, the multiple first frequency domain sequences are merged into a first multi-dimensional frequency domain feature matrix; wherein, the first multi-dimensional frequency domain feature matrix represents the multi-dimensional frequency domain features of the sample time series at the specified time.

[0171] The first multi-dimensional frequency domain feature matrix is ​​input into the LSTM prediction layer of the time series prediction model to predict the second multi-dimensional frequency domain feature matrix; wherein, the second multi-dimensional frequency domain feature matrix represents the multi-dimensional frequency domain features of the target time series at the target time, and it is a merged matrix of multiple second frequency domain sequences of the target time series at the target time.

[0172] In one implementation, the step of inputting the at least one second frequency domain sequence into the frequency-time transform layer of the time series prediction model, and performing inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time series of the target time series includes:

[0173] The second multi-dimensional frequency domain feature matrix is ​​input into the frequency-time transformation layer of the time series prediction model, and the multiple second frequency domain sequences in the second multi-dimensional frequency domain feature matrix are subjected to inverse Fourier transform to obtain multiple second time series; wherein, the multiple second time series include: the overall target time series and the local target time series of the target time series.

[0174] In one implementation, training the time series prediction model based on the loss value of the at least one second time series to obtain the trained time series prediction model includes:

[0175] Based on a preset loss function, the loss value of the target time series is calculated according to the loss values ​​of the at least one second time series;

[0176] Based on the loss value of the target time series, the parameters of the time series prediction model are updated using the backpropagation algorithm.

[0177] In one implementation, the sample time series is a time series of electricity consumption.

[0178] The step of obtaining the sample time series at a specified time includes:

[0179] The power consumption of the target computer system is collected according to the set sampling interval to obtain the time series of the power consumption of the target computer system;

[0180] By setting a time window of a certain length, a sample time series of power consumption with a specified time endpoint is obtained from the time series of power consumption of the target computer system.

[0181] In another specific embodiment, the computer device includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for use in the computer device, and is configured to be executed by one or more processors. The one or more programs include computer-executable instructions for performing the following:

[0182] Get the specified time series at a specified time point to be predicted;

[0183] The specified time series is input into the trained time series prediction model to predict the target time series at the target time.

[0184] The time-series prediction model includes: a time-series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer;

[0185] The time-series feature extraction layer is used to extract at least one first time series from the specified time series; wherein, the at least one first time series includes: the overall time series of the specified time series;

[0186] The time-frequency transformation layer is used to perform Fourier transform on the at least one first time series, transforming the first time series into a first frequency domain sequence of a specified sequence length, to obtain at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0187] The LSTM prediction layer is used to predict at least one second frequency domain sequence of the target time series based on the at least one first frequency domain sequence; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0188] The frequency-time transformation layer is used to perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time sequence of the target time sequence; wherein, the at least one second time sequence corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time sequence includes: the overall target time sequence of the target time sequence;

[0189] The time series prediction model is trained using the time series prediction model training method provided in the foregoing embodiments.

[0190] In one implementation, the specified time series is a time series of the power consumption of the target computer system at a specified time, and the target time series is a time series of the power consumption of the target computer system at a target time.

[0191] The method further includes:

[0192] The power consumption change of the target computer system during the target time period is determined based on the target time series.

[0193] The power distribution of the target computer system is adjusted according to the changes in power consumption of the target computer system during the target time period.

[0194] In one implementation, adjusting the power distribution of the target computer system based on the change in power consumption of the target computer system during the target time period includes:

[0195] When the power consumption of the target computer system during the target time period shows an increasing trend, the capacity of the target computer system is expanded to reduce the power consumption of the target computer system during the target time period.

[0196] When the trend of power consumption of the target computer system during the target time period is decreasing, the target computer system is scaled down to make full use of the power available during the target time period.

[0197] It should be noted that the embodiments concerning computer devices in this application and the embodiments concerning training methods or time series prediction methods for time series prediction models in this application are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the aforementioned implementation of the corresponding training methods or time series prediction methods for time series prediction models, and the repeated parts will not be described again.

[0198] Furthermore, corresponding to the training method and time series prediction method of the time series prediction model provided in the embodiments of this application, based on the same technical concept, the embodiments of this application also provide a storage medium for storing computer-executable instructions;

[0199] In one specific embodiment, the storage medium can be a USB flash drive, optical disc, hard disk, etc., and the computer-executable instructions stored in the storage medium, when executed by a processor, can achieve the following process:

[0200] Obtain the sample time series at a specified time;

[0201] The sample time series is input into the time series feature extraction layer of the time series prediction model to be trained, and at least one first time series is extracted from the sample time series; wherein, the at least one first time series includes: the overall sample time series of the sample time series;

[0202] The at least one first time series is input into the time-frequency transformation layer of the time-series prediction model, and Fourier transform is performed on the at least one first time series to transform the first time series into a first frequency domain sequence of a specified sequence length, thereby obtaining at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0203] The at least one first frequency domain sequence is input into the LSTM prediction layer of the time series prediction model to predict at least one second frequency domain sequence of the target time series at the target time; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0204] The at least one second frequency domain sequence is input into the frequency-time transformation layer of the time series prediction model, and inverse Fourier transform is performed on the at least one second frequency domain sequence to obtain at least one second time series of the target time series; wherein, the at least one second time series corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time series includes: the overall target time series of the target time series;

[0205] The time series prediction model is trained based on the loss value of the at least one second time series to obtain the trained time series prediction model.

[0206] In one implementation, the at least one first time series further includes: a local sample time series of the sample time series;

[0207] The step of extracting at least one first time series from the sample time series includes:

[0208] Extract the overall sample time series from the sample time series;

[0209] One or more local sample time series are extracted from the sample time series using a salient feature extraction network; wherein the local sample time series are the parts of the sample time series that have undergone significant changes.

[0210] In one implementation, the at least one first time series further includes: a local time series of the sample time series;

[0211] The step of extracting at least one first time series from the sample time series includes:

[0212] Extract the overall sample time series from the sample time series;

[0213] One or more local sample time series are extracted from the sample time series using a recent feature extraction network; wherein, the local sample time series is a recent time series in the sample time series that has the same time end point as the overall sample time series.

[0214] In one implementation, performing Fourier transforms on the at least one first time series to transform the first time series into a first frequency domain sequence of a specified length, thereby obtaining at least one first frequency domain sequence, includes:

[0215] A preset number of Fast Fourier Transforms are performed on the multiple first time series to transform them into a preset number of first frequency domain sequences, resulting in multiple first frequency domain sequences; wherein, the multiple first frequency domain sequences correspond one-to-one with the multiple first time series, and any two first frequency domain sequences in the multiple first frequency domain sequences have the same sequence length.

[0216] In one implementation, the step of inputting the at least one first frequency domain sequence into the LSTM prediction layer of the time series prediction model to predict at least one second frequency domain sequence of the target time series at the target time includes:

[0217] Using each first frequency domain sequence as a feature of one dimension, the multiple first frequency domain sequences are merged into a first multi-dimensional frequency domain feature matrix; wherein, the first multi-dimensional frequency domain feature matrix represents the multi-dimensional frequency domain features of the sample time series at the specified time.

[0218] The first multi-dimensional frequency domain feature matrix is ​​input into the LSTM prediction layer of the time series prediction model to predict the second multi-dimensional frequency domain feature matrix; wherein, the second multi-dimensional frequency domain feature matrix represents the multi-dimensional frequency domain features of the target time series at the target time, and it is a merged matrix of multiple second frequency domain sequences of the target time series at the target time.

[0219] In one implementation, the step of inputting the at least one second frequency domain sequence into the frequency-time transform layer of the time series prediction model, and performing inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time series of the target time series includes:

[0220] The second multi-dimensional frequency domain feature matrix is ​​input into the frequency-time transformation layer of the time series prediction model, and the multiple second frequency domain sequences in the second multi-dimensional frequency domain feature matrix are subjected to inverse Fourier transform to obtain multiple second time series; wherein, the multiple second time series include: the overall target time series and the local target time series of the target time series.

[0221] In one implementation, training the time series prediction model based on the loss value of the at least one second time series to obtain the trained time series prediction model includes:

[0222] Based on a preset loss function, the loss value of the target time series is calculated according to the loss values ​​of the at least one second time series;

[0223] Based on the loss value of the target time series, the parameters of the time series prediction model are updated using the backpropagation algorithm.

[0224] In one implementation, the sample time series is a time series of electricity consumption.

[0225] The step of obtaining the sample time series at a specified time includes:

[0226] The power consumption of the target computer system is collected according to the set sampling interval to obtain the time series of the power consumption of the target computer system;

[0227] By setting a time window of a certain length, a sample time series of power consumption with a specified time endpoint is obtained from the time series of power consumption of the target computer system.

[0228] In another specific embodiment, the storage medium can be a USB flash drive, optical disc, hard disk, etc., and the computer-executable instructions stored in the storage medium can achieve the following process when executed by the processor:

[0229] Get the specified time series at a specified time point to be predicted;

[0230] The specified time series is input into the trained time series prediction model to predict the target time series at the target time.

[0231] The time-series prediction model includes: a time-series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer;

[0232] The time-series feature extraction layer is used to extract at least one first time series from the specified time series; wherein, the at least one first time series includes: the overall time series of the specified time series;

[0233] The time-frequency transformation layer is used to perform Fourier transform on the at least one first time series, transforming the first time series into a first frequency domain sequence of a specified sequence length, to obtain at least one first frequency domain sequence; wherein, the specified sequence length is less than the minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds one-to-one with the at least one first time series;

[0234] The LSTM prediction layer is used to predict at least one second frequency domain sequence of the target time series based on the at least one first frequency domain sequence; wherein, the at least one second frequency domain sequence corresponds one-to-one with the at least one first frequency domain sequence;

[0235] The frequency-time transformation layer is used to perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time sequence of the target time sequence; wherein, the at least one second time sequence corresponds one-to-one with the at least one second frequency domain sequence, and the at least one second time sequence includes: the overall target time sequence of the target time sequence;

[0236] The time series prediction model is trained using the time series prediction model training method provided in the foregoing embodiments.

[0237] In one implementation, the specified time series is a time series of the power consumption of the target computer system at a specified time, and the target time series is a time series of the power consumption of the target computer system at a target time.

[0238] The method further includes:

[0239] The power consumption change of the target computer system during the target time period is determined based on the target time series.

[0240] The power distribution of the target computer system is adjusted according to the changes in power consumption of the target computer system during the target time period.

[0241] In one implementation, adjusting the power distribution of the target computer system based on the change in power consumption of the target computer system during the target time period includes:

[0242] When the power consumption of the target computer system during the target time period shows an increasing trend, the capacity of the target computer system is expanded to reduce the power consumption of the target computer system during the target time period.

[0243] When the trend of power consumption of the target computer system during the target time period is decreasing, the target computer system is scaled down to make full use of the power available during the target time period.

[0244] It should be noted that the embodiments concerning storage media in this application and the embodiments concerning the training method or timing prediction method of the timing prediction model in this application are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the corresponding training method or timing prediction method of the timing prediction model mentioned above, and the repeated parts will not be described again.

[0245] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0246] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0247] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0248] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0249] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0250] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0251] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0252] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0253] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0254] The embodiments of this application can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. One or more embodiments of this application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can reside in local and remote computer storage media, including storage devices.

[0255] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0256] The above description is merely an embodiment of this document and is not intended to limit the scope of this document. Various modifications and variations can be made to this document by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this document should be included within the scope of the claims of this document.

Claims

1. A method for training a time series prediction model, the method comprising: obtaining a sample time series at a specified time point; wherein the sample time series is a time series of power consumption of a target computer system at the specified time point; inputting the sample time series into a time series feature extraction layer of a time series prediction model to be trained, and extracting at least one first time series from the sample time series; wherein the at least one first time series comprises an overall time series of the sample time series and a local time series in the sample time series; inputting the at least one first time series into a time-frequency transformation layer of the time series prediction model, and performing Fourier transform on the at least one first time series respectively to transform the at least one first time series into a first frequency domain sequence of a specified sequence length respectively, to obtain at least one first frequency domain sequence; wherein the specified sequence length is smaller than a minimum sequence length of the at least one first time series, and the at least one first frequency domain sequence corresponds to the at least one first time series one by one; inputting the at least one first frequency domain sequence into an LSTM prediction layer of the time series prediction model, and predicting at least one second frequency domain sequence of a target time series at a target time point; wherein the at least one second frequency domain sequence corresponds to the at least one first frequency domain sequence one by one, and the target time series is a time series of power consumption of the target computer system at the target time point; inputting the at least one second frequency domain sequence into a frequency-time transformation layer of the time series prediction model, and performing inverse Fourier transform on the at least one second frequency domain sequence respectively to obtain at least one second time series of the target time series; wherein the at least one second time series corresponds to the at least one second frequency domain sequence one by one, and the at least one second time series comprises an overall target time series; training the time series prediction model based on a loss value of the at least one second time series, to obtain a trained time series prediction model, which is a prediction model of power consumption of the target computer system; wherein the inputting the at least one first frequency domain sequence into the LSTM prediction layer of the time series prediction model to predict the at least one second frequency domain sequence of the target time series at the target time point comprises: taking each first frequency domain sequence as a dimension of a feature, and merging the plurality of first frequency domain sequences into a first multi-dimensional frequency domain feature matrix; wherein the first multi-dimensional frequency domain feature matrix represents multi-dimensional frequency domain features of the sample time series; inputting the first multi-dimensional frequency domain feature matrix into the LSTM prediction layer of the time series prediction model to predict a second multi-dimensional frequency domain feature matrix; wherein the second multi-dimensional frequency domain feature matrix represents multi-dimensional frequency domain features of the target time series, and is a merged matrix of a plurality of second frequency domain sequences of the target time series at the target time point.

2. The method of claim 1, wherein the extracting at least one first time series from the sample time series comprises: ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ extracting an overall time sequence from the sample time sequence by an overall feature extraction network; extracting one or more local time sequences from the sample time sequence by a salient feature extraction network; wherein the local time sequence is a salient time sequence in which salient change occurs in the sample time sequence.

3. The training method of the time series prediction model according to claim 1, wherein, the extracting at least one first time sequence from the sample time sequence comprises: extracting an overall time sequence from the sample time sequence by an overall feature extraction network; extracting one or more local time sequences from the sample time sequence by a recent feature extraction network; wherein the local time sequence is a recent time sequence with the same time end point as the overall time sequence in the sample time sequence.

4. The method of training a time series forecasting model according to claim 2 or 3, wherein, the Fourier transforming the at least one first time sequence respectively to transform the first time sequence into a first frequency domain sequence with a specified sequence length, to obtain at least one first frequency domain sequence, comprises: performing fast Fourier transform of a preset point number on the plurality of first time sequences respectively to transform the plurality of first time sequences into first frequency domain sequences with a preset point number respectively, to obtain a plurality of first frequency domain sequences; wherein the plurality of first frequency domain sequences correspond to the plurality of first time sequences one by one, and any two first frequency domain sequences in the plurality of first frequency domain sequences have the same sequence length.

5. The training method according to claim 1, the inputting the at least one second frequency domain sequence into a frequency-time conversion layer of the time series prediction model to perform inverse Fourier transform on the at least one second frequency domain sequence respectively to obtain at least one second time sequence of the target time sequence, comprises: inputting the second multi-dimensional frequency domain feature matrix into the frequency-time conversion layer of the time series prediction model to perform inverse Fourier transform on a plurality of second frequency domain sequences in the second multi-dimensional frequency domain feature matrix respectively to obtain a plurality of second time sequences; wherein the plurality of second time sequences comprise an overall target time sequence and a local target time sequence.

6. The training method of claim 1, wherein, the training the time series prediction model based on the loss value of the at least one second time sequence to obtain a trained time series prediction model, comprises: calculating the loss value of the target time sequence based on the loss value of the at least one second time sequence according to a preset loss function; updating parameters of the time series prediction model by using a back propagation algorithm based on the loss value of the target time sequence. 7.The method of Claim 1, wherein, the sample time sequence is a time sequence of power consumption; the obtaining the sample time sequence of power consumption of the target computer system at a specified time point comprises: collecting power consumption of the target computer system at a set sampling interval to obtain a time sequence of power consumption of the target computer system; obtaining the sample time sequence of power consumption with a time end point of a specified time point from the time sequence of power consumption of the target computer system by a time window with a set length.

8. A time series prediction method, the method comprising: obtaining a specified time sequence of a specified time point to be predicted, the specified time sequence being a time sequence of power consumption of a target computer system at the specified time point; inputting the specified time sequence into a trained time series prediction model to predict a target time sequence of a target time point, the target time sequence being a time sequence of power consumption of the target computer system at the target time point; wherein the time series prediction model comprises a time series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer; the time series feature extraction layer is configured to extract at least one first time sequence from the specified time sequence, wherein the at least one first time sequence comprises an overall time sequence and a local time sequence; the time-frequency transformation layer is configured to perform Fourier transform on the at least one first time sequence to transform the first time sequence into a first frequency domain sequence of a specified sequence length, to obtain at least one first frequency domain sequence, wherein the specified sequence length is less than a minimum sequence length of the at least one first time sequence, and the at least one first frequency domain sequence corresponds to the at least one first time sequence in a one-to-one manner; the LSTM prediction layer is configured to predict at least one second frequency domain sequence of the target time sequence based on the at least one first frequency domain sequence, wherein the at least one second frequency domain sequence corresponds to the at least one first frequency domain sequence in a one-to-one manner; the frequency-time transformation layer is configured to perform inverse Fourier transform on the at least one second frequency domain sequence to obtain at least one second time sequence of the target time sequence, wherein the at least one second time sequence corresponds to the at least one second frequency domain sequence in a one-to-one manner, and the at least one second time sequence comprises an overall target time sequence; wherein the time series prediction model is trained by the training method of the time series prediction model according to any one of claims 1-7.

9. The time series prediction method of claim 8, wherein the method further comprises: determining a power consumption change of the target computer system in a target time period according to the target time sequence; adjusting power distribution of the target computer system according to the power consumption change of the target computer system in the target time period.

10. The time series prediction method of claim 9, wherein the adjusting power distribution of the target computer system according to the power consumption change of the target computer system in the target time period comprises: when a trend of the power consumption change of the target computer system in the target time period is increasing, expanding the target computer system to reduce power consumption of the target computer system in the target time period; when the trend of the power consumption change of the target computer system in the target time period is decreasing, shrinking the target computer system to increase power consumption of the target computer system in the target time period.

11. A training device of a time series prediction model, the device comprising: a sample obtaining module, configured to obtain a sample time sequence at a specified time; wherein the sample time sequence is a time sequence of power consumption of a target computer system at the specified time; a feature extraction module, configured to input the sample time sequence into a time sequence feature extraction layer of a time series prediction model to be trained, and extract at least one first time sequence from the sample time sequence; wherein the at least one first time sequence includes an overall time sequence of the sample time sequence and a local time sequence in the sample time sequence; a time-frequency transformation module, configured to input the at least one first time sequence into a time-frequency transformation layer of the time series prediction model, and perform Fourier transform on the at least one first time sequence respectively, so as to transform the first time sequence into a first frequency domain sequence of a specified sequence length, and obtain at least one first frequency domain sequence; wherein the specified sequence length is smaller than a minimum sequence length of the at least one first time sequence, and the at least one first frequency domain sequence corresponds to the at least one first time sequence one by one; a prediction module, configured to input the at least one first frequency domain sequence into an LSTM prediction layer of the time series prediction model, and predict at least one second frequency domain sequence of a target time sequence at a target time; wherein the at least one second frequency domain sequence corresponds to the at least one first frequency domain sequence one by one; the target time sequence is a time sequence of power consumption of the target computer system at the target time; a frequency-time transformation module, configured to input the at least one second frequency domain sequence into a frequency-time transformation layer of the time series prediction model, and perform inverse Fourier transform on the at least one second frequency domain sequence respectively, so as to obtain at least one second time sequence of the target time sequence; wherein the at least one second time sequence corresponds to the at least one second frequency domain sequence one by one, and the at least one second time sequence includes an overall target time sequence; a training module, configured to train the time series prediction model based on a loss value of the at least one second time sequence, so as to obtain a trained time series prediction model, which is a prediction model of power consumption of the target computer system; wherein the inputting of the at least one first frequency domain sequence into the LSTM prediction layer of the time series prediction model to predict the at least one second frequency domain sequence of the target time sequence at the target time includes: combining the plurality of first frequency domain sequences into a first multi-dimensional frequency domain feature matrix as a feature of each first frequency domain sequence; wherein the first multi-dimensional frequency domain feature matrix represents multi-dimensional frequency domain features of the sample time sequence; inputting the first multi-dimensional frequency domain feature matrix into the LSTM prediction layer of the time series prediction model to predict a second multi-dimensional frequency domain feature matrix; wherein the second multi-dimensional frequency domain feature matrix represents multi-dimensional frequency domain features of the target time sequence, which is a combined matrix of a plurality of second frequency domain sequences of the target time sequence at the target time.

12. A time series prediction device, the device comprising: The sequence obtaining module is configured to obtain a specified time sequence at a specified time point to be predicted, the specified time sequence being a time sequence of power consumption of a target computer system at the specified time point. The prediction module is configured to input the specified time sequence into a trained time series prediction model, and predict a target time sequence at a target time point, the target time sequence being a time sequence of power consumption of the target computer system at the target time point. The time series prediction model comprises a time series feature extraction layer, a time-frequency transformation layer, an LSTM prediction layer, and a frequency-time transformation layer. The time series feature extraction layer is configured to extract at least one first time sequence from the specified time sequence, wherein the at least one first time sequence comprises an overall time sequence and a local time sequence. The time-frequency transformation layer is configured to perform Fourier transform on the at least one first time sequence respectively, and transform the first time sequence into a first frequency domain sequence of a specified sequence length, to obtain at least one first frequency domain sequence, wherein the specified sequence length is smaller than a minimum sequence length of the at least one first time sequence, and the at least one first frequency domain sequence corresponds to the at least one first time sequence one by one. The LSTM prediction layer is configured to predict at least one second frequency domain sequence of the target time sequence at the target time point based on the at least one first frequency domain sequence, wherein the at least one second frequency domain sequence corresponds to the at least one first frequency domain sequence one by one. The frequency-time transformation layer is configured to perform inverse Fourier transform on the at least one second frequency domain sequence respectively, to obtain at least one second time sequence of the target time sequence, wherein the at least one second time sequence corresponds to the at least one second frequency domain sequence one by one, and the at least one second time sequence comprises an overall target time sequence. The time series prediction model is trained by the training method of the time series prediction model according to any one of claims 1-7.

13. A computer device, comprising: The device comprises: a processor; and a memory arranged to store computer executable instructions configured to be executed by the processor, the executable instructions comprising instructions for performing steps in the method according to any one of claims 1-7 or according to any one of claims 8-10.

14. A storage medium, characterized by The storage medium is configured to store computer executable instructions, the executable instructions causing a computer to perform the method according to any one of claims 1-7 or according to any one of claims 8-10.

Images