System for prediction, method and device for training, and computer readable storage medium
By combining channel joint prediction and channel independent prediction into a system, and using a refiner model to optimize the spliced sequence, the problem of insufficient prediction accuracy in existing methods is solved, and more efficient time series prediction results are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-30
- Publication Date
- 2026-03-31
AI Technical Summary
Existing independent channel forecasting and joint channel forecasting methods each have their shortcomings in multivariate time series forecasting, and fusion methods do not significantly improve forecasting accuracy.
A novel fusion system is adopted, which first predicts multivariate sequences through a channel joint predictor and a refiner model, then optimizes the prediction results of each channel, and uses the refiner model to optimize the spliced sequence, thereby improving the prediction accuracy.
It improves the accuracy of time series forecasting, and experimental results show that it significantly enhances the performance of various basic forecasting models on a variety of real-world datasets.
Smart Images
Figure CN121765210A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of forecasting, and more specifically to forecasting time series data. Background Technology
[0002] A time series is a sequence of data points arranged in chronological order. Typically, the time intervals in a time series can be constant, allowing it to be analyzed as discrete data. Time series analysis is widely used in various fields, such as power systems, transportation systems, finance, weather forecasting, and biomedicine.
[0003] Time series analysis is typically divided into four tasks: prediction, imputation, anomaly detection, and classification. Prediction aims to predict future values or trends from an input series. Imputation aims to recover missing data. Anomaly detection aims to identify anomalies or abnormal patterns in the series. Classification aims to assign labels or categories to a given time series and is widely used in medical diagnostics.
[0004] In the prediction of multivariate time series, existing methods can be divided into two types based on their multivariate modeling approaches. One is channel-independent prediction (CI), such as PatchTST, which divides the multivariate series into multiple univariate predictions and models each variable separately. The other is channel-dependent prediction (CD), such as Informer and FEDformer, which models multiple variables together by considering the relationships between them.
[0005] Figure 1 The diagram illustrates independent channel prediction and joint channel prediction. For example... Figure 1 As shown, assuming the multivariate sequence is divided into three variables, in channel joint prediction, the model f is applied to all three variables together. θ In channel-independent prediction, f is modeled separately for each of the three variables. θ1 f θ2 f θ3 .
[0006] Independent channel prediction and joint channel prediction each have their own characteristics. Independent channel prediction considers only a single variable, making the model simpler, but it loses some key information because it doesn't account for the relationships between different sequences. Joint channel prediction considers more comprehensive information, but the model is also more complex.
[0007] There are also fusion methods that combine channel-independent prediction and channel-joint prediction, attempting to combine the advantages of both. Such fusion methods include C-LoRA, which performs channel-independent processing at the input layer, models each channel separately using LoRA, and then performs joint channel prediction. However, these fusion methods do not provide satisfactory improvements in model performance. Summary of the Invention
[0008] A brief overview of this disclosure is provided below to provide a basic understanding of certain aspects of it. It should be understood that this overview is not an exhaustive summary of the disclosure. It is not intended to identify key or essential parts of the disclosure, nor is it intended to limit its scope. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.
[0009] According to one aspect of this disclosure, a system for prediction is provided, the system comprising: a predictor configured to predict M first sequences input and output M second sequences corresponding to the M first sequences, wherein the M first sequences are sequences within a first time period and the M second sequences are sequences within a second time period following the first time period; and a refiner model comprising M refiners and configured to concatenate each of the input M first sequences with a corresponding sequence from the M second sequences output by the predictor, optimize the sequences belonging to the M second sequences in the concatenated M sequences based on each of the concatenated M sequences using the M refiners, and output optimized M second sequences, wherein M is an integer greater than or equal to 1, and wherein the system is trained such that the optimized M second sequences approximate the true sequences within the second time period.
[0010] According to a preferred embodiment, the predictor is a channel joint predictor and is configured to jointly predict M first sequences, which are correlated with each other.
[0011] According to a preferred embodiment, the predictor is an M-channel independent predictor and is configured to predict the corresponding sequences in the M first sequences individually.
[0012] According to a preferred embodiment, each of the M refiners is configured to append one of the M first sequences before the corresponding sequence in the M second sequences.
[0013] According to a preferred embodiment, each of the M refiners is configured to append one of the M first sequences to the corresponding sequence in the M second sequences.
[0014] According to a preferred embodiment, the M first sequences, the M second sequences, and the optimized M second sequences are sequences arranged in chronological order.
[0015] According to a preferred embodiment, the predictor can be a transformer-based model, such as Informer, FEDformer, or iTranformer, or a linear model, such as DLinear, TSMixer, or FreTS.
[0016] According to a preferred embodiment, each of the M refiners includes a layer of linear neural network.
[0017] According to a preferred embodiment, the M first sequences and M second sequences are normalized before being input into the refiner model, and the refiner model is denormalized before outputting the optimized M second sequences.
[0018] According to a preferred embodiment, the training system includes first training a predictor, and then training a refiner model while keeping the parameters of the predictor unchanged.
[0019] According to a preferred embodiment, the predictor is trained by minimizing the mean square error or absolute error between the M second sequences and the true sequence.
[0020] According to a preferred embodiment, the refiner model is trained by minimizing the mean square error or absolute error between the optimized M second sequences and the true sequence.
[0021] According to another aspect of this disclosure, a method for training the aforementioned prediction system is provided, comprising: training a predictor such that M second sequences output by the predictor approximate the true sequences; and after training the predictor, and while keeping the parameters of the predictor unchanged, training a refiner model such that the optimized M second sequences output by the M refiners approximate the true sequences.
[0022] According to a preferred embodiment, the method further includes training the predictor to minimize the mean square error or absolute error between the M second sequences and the true sequence.
[0023] According to a preferred embodiment, the method further includes training the refiner model to minimize the mean square error or absolute error between the optimized M second sequences and the true sequence.
[0024] According to a preferred embodiment, the method further includes aligning and normalizing the M first sequences and M second sequences before inputting them into the refiner model, and denormalizing them before the refiner model outputs optimized M second sequences.
[0025] According to another aspect of this disclosure, an apparatus for training the aforementioned prediction system is provided, comprising: a first training device configured to train a predictor such that M second sequences output by the predictor approximate true sequences; and a second training device configured to, after training the predictor and while keeping the parameters of the predictor unchanged, train a refiner model such that the optimized M second sequences output by the refiner model approximate true sequences.
[0026] According to a preferred embodiment, the first training device is further configured to train a predictor such that the mean square error or absolute error between the M second sequences and the true sequence is minimized.
[0027] According to a preferred embodiment, the second training device is further configured to train a refiner model such that the mean square error or absolute error between the optimized M second sequences and the true sequence is minimized.
[0028] According to another aspect of this disclosure, a computer-readable storage medium storing a program is provided, which, when executed by a processor, causes the computer to implement the above-described training method.
[0029] In accordance with other aspects of this disclosure, corresponding computer program code and computer program products are also provided.
[0030] The systems for prediction, methods for training, and apparatus disclosed herein enable effective improvement in the accuracy of time series predictions.
[0031] These and other advantages of this disclosure will become more apparent from the following detailed description of preferred embodiments of this disclosure in conjunction with the accompanying drawings. Attached Figure Description
[0032] To further illustrate the above and other advantages and features of this disclosure, the specific embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. These drawings, together with the following detailed description, are included in and form a part of this specification. Elements having the same function and structure are indicated by the same reference numerals. It should be understood that these drawings only depict typical examples of this disclosure and should not be considered as limiting the scope of this disclosure. In the drawings:
[0033] Figure 1 The diagram illustrates joint channel forecasting and independent channel forecasting.
[0034] Figure 2 A system for prediction according to an embodiment of the present disclosure is illustrated schematically;
[0035] Figure 3 schematically shown Figure 2 The operation of the system;
[0036] Figure 4 schematically shown Figure 2 The operation of the refiner;
[0037] Figure 5 This is a flowchart of a method for training according to an embodiment of the present disclosure;
[0038] Figure 6 This is a block diagram of a training device according to an embodiment of the present disclosure; and
[0039] Figure 7 The improved prediction results achieved by the prediction system according to the embodiments of this disclosure are visually illustrated using a graph.
[0040] Figure 8 This is a block diagram of an exemplary structure of a general-purpose personal computer in which methods and / or devices according to embodiments of this disclosure can be implemented. Detailed Implementation
[0041] Exemplary embodiments of this disclosure will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of actual embodiments are described in the specification. However, it should be understood that many implementation-specific decisions must be made in the development of any such actual embodiment to achieve the developer's specific goals, such as complying with constraints related to the system and business, and these constraints may vary depending on the implementation. Furthermore, it should be understood that while development work can be very complex and time-consuming, such development work is merely a routine task for those skilled in the art who benefit from this disclosure.
[0042] It should also be noted that, in order to avoid obscuring this disclosure with unnecessary details, only the equipment structure and / or processing steps closely related to the solution according to this disclosure are shown in the accompanying drawings, while other details that are not closely related to this disclosure are omitted.
[0043] To address the problems of existing technologies, this disclosure proposes a novel system that integrates joint channel prediction and independent channel prediction. The system inputs historical time series data into a joint channel predictor. Based on the output time series of the joint predictor, a refiner is built for each channel. The output time series and the historical time series are concatenated together and used as input to the refiner, which then improves the prediction results for each channel. The refiner can be, for example, a simple linear neural network layer for each channel. Linear layers, also known as fully connected layers, are a standard layer in neural networks where each input node is connected to each output node, and each node in a linear layer has an associated weight vector that is used to compute the output value.
[0044] The following is combined Figures 2 to 4 A system 100 for prediction according to an embodiment of the present disclosure is described in detail.
[0045] like Figure 2 As shown, system 100 includes a predictor 10 and a refiner model 20. First, M first sequences 11 are input to the predictor 10. The M first sequences 11 are, for example,... Figure 3 sequence in Where L represents the sequence length, i.e., the time length of the sequence, and M represents the number of variables (channels). Predictor 10 predicts the changing trends of M first sequences 11 over a future time period T and outputs M second sequences 12. The M second sequences 12 are, for example, Figure 3 sequence in Where T represents the length of a period of time following time period L.
[0046] Next, M first sequences 11 and M second sequences 12 are input into the refiner model 20. The refiner model 20 contains M refiners f1, ..., f2. M Each of the M first sequences 11 is concatenated with its corresponding sequence from the M second sequences 12 to obtain M concatenated sequences. It should be understood that the second sequences 12 can be concatenated before or after the first sequences 11.
[0047] Next, M refiners f1, ..., f M Based on the M concatenated sequences, the sequences belonging to the M second sequences 12 within each of the M concatenated sequences are optimized to obtain M optimized second sequences 12'. For example, the optimized second sequence 12' is... Figure 3 sequence in
[0048] According to a preferred embodiment, the M first sequences 11 and M second sequences 12 can be normalized before being input into the refiner model 20, that is, the data can be mapped to the range of 0 to 1, and correspondingly, they can be denormalized before the refiner model 20 outputs the optimized M second sequences 12'. This can reduce the adverse effects caused by inconsistent data distribution, improve the system's generalization performance, and accelerate training.
[0049] It should be noted that the predictor 10 can be a transformer-based model, such as Informer, FEDformer or iTranformer, or a linear model, such as DLinear, TSMixer or FreTS.
[0050] It should also be noted that refiners 1 to M can be implemented using a single-layer linear neural network for the concatenated sequences of the M channels, such as... Figure 4 As shown. However, it should be understood that the implementation of refiners 1 to M is not limited to this, but can also be implemented using convolutional neural networks, recurrent neural networks, transformers, etc.
[0051] It should also be noted that the training of system 100 can be performed simultaneously on predictor 10 and refiner model 20.
[0052] According to a preferred embodiment, the training of system 100 can be divided into two stages. First, predictor 10 is trained so that the predicted future sequence of the training set data is as close as possible to the true future sequence. Then, after training predictor 10, refiner model 20 is trained with the parameters of predictor 10 fixed, so that the predicted future sequence of the training set data is as close as possible to the true future sequence.
[0053] It should be understood that the training of the refiner model 20 includes training refiners 1 to M separately, and the parameters of the trained refiners 1 to M are either the same or different from each other.
[0054] In the case of two-stage training, assuming that predictor 10 is a channel joint predictor, the objective function can be set, for example, as the mean squared error loss function l′ as shown in Equation 1).
[0055]
[0056] Where N represents the number of training samples, θ represents the parameters of predictor 10, and y i x represents the true sequence in the training set data. i Let f represent the input sequence of the i-th training sample. θ (x i ) represents the prediction result of predictor 10 for the i-th training sample.
[0057] The objective function of the refiner model 20 is set as the mean squared error loss function l as shown in Equation 2).
[0058]
[0059] Where N represents the number of training samples, M represents the number of channels, and θ1, ..., θ c This indicates the parameters for each channel in Refiner Model 20. This represents the c-th sequence out of the M true sequences in the training set data. f represents the c-th element in the M input sequences of the i-th training sample. θ (x i ) c This represents the output of predictor 10 of the M predicted sequences for the i-th training sample, which is the c-th sequence. This represents the prediction result of the refiner model 20 for the c-th channel.
[0060] In the two-stage training, the loss function l′ is first used to train the predictor 10, and after the training of the predictor 10 is completed, the loss function l is used to train the refiner model 20 while keeping the parameters of the predictor 10 unchanged.
[0061] It should be noted that the construction of loss functions l′ and l as shown in Equations 1) and 2) above is merely an example, and this disclosure is not limited thereto. Rather, any other suitable method may be used, as long as the prediction results of predictor 10 and refiner model 20 are as close as possible to the true values.
[0062] When the predictor 10 and the refiner model 20 are trained simultaneously, the total loss function L can be, for example, the sum of equations 1) and 2) above, as shown in equation 3) below.
[0063] L= αl′+ βl (3)
[0064] Here, α and β represent weights and can be set according to actual needs.
[0065] It should be understood that although the loss function l′ in Equation 1) above is constructed based on joint channel prediction, the loss function l′ can also be constructed similarly based on independent channel prediction. For the sake of brevity, the construction of the loss function based on independent channel prediction will not be described in detail here.
[0066] Figure 5 This is a flowchart of a training method 500 based on this disclosure. The following is in conjunction with… Figure 2 The system 100 describes the method 500 for graphs.
[0067] First, in step 501, a predictor is trained so that the M second sequences output by the predictor are close to the real sequences. For example, in this embodiment, the predictor 10 is trained using a loss function l′ as shown in Equation 1), so that the prediction results output by the predictor 10 are close to the real sequences in the dataset.
[0068] Next, in step 502, after training the predictor and keeping the predictor parameters unchanged, the refiner model is trained so that the optimized M second sequences output by the refiner model are close to the real sequences. For example, in this embodiment, the refiner model 20 is trained using a loss function l as shown in Equation 2), so that the prediction results output by the refiner model 20 are close to the real sequences in the dataset.
[0069] The methods discussed above can be implemented entirely by a computer-executable program, or partially or entirely using hardware and / or firmware. When implemented in hardware and / or firmware, or when a computer-executable program is loaded into a hardware device capable of running the program, the training apparatus described below is realized. Hereinafter, an overview of these apparatuses is given without repeating some details already discussed above; however, it should be noted that while these apparatuses can perform the methods described above, the methods may not necessarily employ or be performed by those components of the described apparatus.
[0070] Figure 6 This is a block diagram of a training device 600 according to the present disclosure. Device 600 includes a first training unit 601 and a second training unit 602. The first training unit 601 is configured to train a predictor such that the M second sequences output by the predictor approximate the true sequences. The second training unit 602 is configured to train a refiner model after training the predictor, while keeping the predictor parameters unchanged, such that the optimized M second sequences output by the refiner model approximate the true sequences.
[0071] According to a preferred embodiment, the first training device 601 is further configured to train the predictor to minimize the mean square error or absolute error between the M second sequences and the true sequence.
[0072] According to a preferred embodiment, the second training device 602 is further configured to train the refiner model such that the mean square error or absolute error between the optimized M second sequences and the true sequence is minimized.
[0073] Figure 6 The device 600 shown for training corresponds to Figure 5 The training method 500 is shown. Therefore, the relevant details of the various devices in the training device 600 have already been described in [the following text is missing]. Figure 5The training method is described in detail in the description of 500, and will not be repeated here.
[0074] experiment
[0075] The table below compares the prediction results of the system used for prediction based on this disclosure with those of traditional models. Experiments were conducted on the datasets ETTh1, ETTm1, and Electricity, and the models included the linear model DLinear, and deep models based on transformer architectures iTransformer, FEDformer, and Informer. Mean squared error (MSE) and absolute error (MAE) are commonly used as evaluation metrics for time series prediction; lower values indicate better performance.
[0076]
[0077] Table 1
[0078] As shown in Table 1, the prediction system based on this disclosure can effectively improve prediction accuracy.
[0079] Figure 7 The graph visually demonstrates how the predictions output by the system used for prediction according to this disclosure more closely match the actual sequences. For example... Figure 7 As shown, the optimized prediction result (curve c) is closer to the true value of the sequence (curve a) than the prediction result of the traditional model (curve b).
[0080] Therefore, the prediction system 100 and training method 500 of this disclosure, which combine channel joint and channel independent strategies in a single network, enable the combination of the advantages of both channel strategies to improve the accuracy of time series prediction.
[0081] In addition, experiments have verified that the refiner implemented according to this disclosure can improve the performance of a variety of underlying prediction models on various real-world datasets.
[0082] Each component module and unit in the aforementioned device or equipment can be configured via software, firmware, hardware, or a combination thereof. Specific means or methods of configuration are well known to those skilled in the art and will not be elaborated upon here. When implemented via software or firmware, data can be transferred from a storage medium or network to a computer with a dedicated hardware architecture (e.g., [missing information]). Figure 8 The general-purpose computer 800 shown is equipped with the programs that constitute the software, and when various programs are installed, the computer is able to perform various functions, etc.
[0083] Figure 8This is a block diagram illustrating an exemplary structure of a general-purpose personal computer in which the methods and / or apparatuses according to embodiments of the present invention can be implemented. For example... Figure 8 As shown, the Central Processing Unit (CPU) 801 performs various processes according to the program stored in the Read-Only Memory (ROM) 802 or the program loaded into the Random Access Memory (RAM) 803 from the storage device 808. The RAM 803 also stores data required as needed when the CPU 801 performs various processes, etc. The CPU 801, ROM 802, and RAM 803 are connected to each other via a bus 804. An input / output interface 805 is also connected to the bus 804.
[0084] The following components are connected to input / output interface 805: input device 806 (including keyboard, mouse, etc.), output device 807 (including display, such as cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.), storage device 808 (including hard disk, etc.), and communication device 809 (including network interface card, such as LAN card, modem, etc.). Communication device 809 performs communication processing via a network such as the Internet. Drive 810 may also be connected to input / output interface 805 as needed. Removable media 811, such as disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 810 as needed, so that computer programs read from them can be installed into storage device 808 as needed.
[0085] When the above series of processes are implemented through software, the program constituting the software is installed from a network such as the Internet or a storage medium such as a removable medium 811.
[0086] Those skilled in the art will understand that such storage media are not limited to Figure 8 The illustration shows a removable medium 811 containing a program, distributed separately from the device to provide the program to the user. Examples of removable media 811 include disks (including floppy disks (registered trademark)), optical disks (including optical disc read-only memory (CD-ROM) and digital versatile disks (DVD)), magneto-optical disks (including mini-discs (MD (registered trademark))), and semiconductor memory. Alternatively, the storage medium may be ROM 802, a hard disk included in storage device 808, etc., containing programs and distributed to the user along with the device containing them.
[0087] This disclosure also provides corresponding computer program code and a computer program product storing machine-readable instruction code. When the instruction code is read and executed by a machine, the training method 500 described above according to the embodiments of this disclosure can be performed.
[0088] Accordingly, the storage medium configured to carry the aforementioned program product storing machine-readable instruction code is also included in this disclosure. The storage medium includes, but is not limited to, floppy disks, optical disks, magneto-optical disks, memory cards, memory sticks, etc.
[0089] Based on the above description, the embodiments of this disclosure provide the following technical solutions, but are not limited thereto.
[0090] Option 1. A system for prediction, the system comprising:
[0091] A predictor is configured to predict M first sequences given as input and output M second sequences corresponding to the M first sequences, wherein the M first sequences are sequences within a first time interval, and the M second sequences are sequences within a second time interval following the first time interval; and
[0092] A refiner model includes M refiners configured to concatenate each of the M input first sequences with the corresponding sequence from the M second sequences output from the predictor. Based on each of the concatenated M sequences, the M refiners are used to optimize the sequences belonging to the M second sequences within the concatenated M sequences, and the optimized M second sequences are output.
[0093] Where M is an integer greater than or equal to 1, and
[0094] The system is trained such that the optimized M second sequences approximate the true sequences within the second time period.
[0095] Option 2. The system according to Option 1, wherein the predictor is a channel joint predictor and is configured to jointly predict M first sequences, the M first sequences being correlated with each other.
[0096] Scheme 3. Based on the system of Scheme 1, wherein the predictor is an M-channel independent predictor and is configured to predict the corresponding sequence in each of the M first sequences individually.
[0097] Option 4. A system based on any one of Options 1 to 3, wherein each of the M refiners is configured to append one of the M first sequences to the corresponding sequence in the M second sequences.
[0098] Option 5. A system according to any one of Options 1 to 3, wherein each of the M refiners is configured to append one of the M first sequences to the corresponding sequence in the M second sequences.
[0099] Scheme 6. A system based on any one of Schemes 1 to 3, wherein the M first sequences, the M second sequences, and the optimized M second sequences are sequences arranged in chronological order.
[0100] Option 7. A system based on any one of Options 1 to 3, wherein the predictor is a transformer-based model such as Informer, FEDformer, or iTranformer, or a linear model such as DLinear, TSMixer, or FreTS.
[0101] Option 8. A system based on any one of Options 1 to 3, wherein each of the M refiners comprises a layer of linear neural network.
[0102] Scheme 9. A system based on any one of Schemes 1 to 3, wherein the M first sequences and M second sequences are normalized before being input into the refiner model, and the refiner model is denormalized before outputting the optimized M second sequences.
[0103] Option 10. A system according to any one of Options 1 to 3, wherein the training system includes first training a predictor, and then training a refiner model while keeping the parameters of the predictor unchanged.
[0104] Scheme 11. The system according to Scheme 10, wherein the predictor is trained by minimizing the mean square error or absolute error between the M second sequences and the true sequence.
[0105] Scheme 12. The system according to Scheme 10, wherein M refiners are trained by minimizing the mean square error or absolute error between the optimized M second sequences and the true sequence.
[0106] Option 13. A method for training a system according to any one of Options 1 to 8, comprising:
[0107] Train the predictor so that the M second sequences output by the predictor closely approximate the true sequences; and
[0108] After training the predictor, and with the predictor parameters remaining unchanged, the refiner model is trained so that the optimized M second sequences output by the refiner model are close to the true sequences.
[0109] Scheme 14. According to the method of Scheme 13, the predictor is further trained to minimize the mean square error or absolute error between the M second sequences and the true sequence.
[0110] Option 15. According to the method of Option 13, it further includes training the refiner model to minimize the mean square error or absolute error between the optimized M second sequences and the true sequence.
[0111] Option 16. The method according to any one of Options 13 to 15 further includes normalizing the M first sequences and M second sequences before inputting them into the refiner model, and denormalizing them before outputting the optimized M second sequences from the refiner model.
[0112] Option 17. An apparatus for training a system according to any one of Options 1 to 8, comprising:
[0113] A first training device is configured to train a predictor such that the M second sequences output by the predictor approximate the true sequences; and
[0114] The second training device is configured to train the refiner model after training the predictor, while keeping the parameters of the predictor unchanged, such that the optimized M second sequences output by the refiner model are close to the true sequences.
[0115] Option 18. The apparatus according to Option 17, wherein the first training device is configured to train the predictor to minimize the mean square error or absolute error between the M second sequences and the true sequence.
[0116] Option 19. The apparatus according to Option 17, wherein the second training device is configured to train the refiner model such that the mean square error or absolute error between the optimized M second sequences and the true sequence is minimized.
[0117] Scheme 20. A computer-readable storage medium storing a program that, when executed by a processor, causes the computer to implement the method according to any one of Schemes 13 to 16.
[0118] Finally, it should 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. Furthermore, unless otherwise specified, 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.
[0119] While the embodiments of this disclosure have been described in detail above with reference to the accompanying drawings, it should be understood that the embodiments described above are merely illustrative of this disclosure and do not constitute a limitation thereof. Those skilled in the art can make various modifications and alterations to the above embodiments without departing from the spirit and scope of this disclosure. Therefore, the scope of this disclosure is defined only by the appended claims and their equivalents.
Claims
1. A system for prediction, the system comprising: a predictor configured to make predictions for M input first sequences, and output M second sequences respectively corresponding to the M first sequences, wherein the M first sequences are sequences within a first time period, and the M second sequences are sequences within a second time period after the first time period; and a refiner model comprising M refiners, and configured to splice each of the M input first sequences with a corresponding one of the M second sequences output from the predictor, optimize sequences of the M second sequences with the M refiners respectively based on each of the spliced M sequences, and output optimized M second sequences, wherein M is an integer greater than or equal to 1, and wherein the system is trained such that the optimized M second sequences approximate true sequences within the second time period.
2. The system of claim 1, wherein, the predictor is a channel joint predictor, and is configured to make joint predictions on the M first sequences, which have correlations among each other.
3. The system of claim 1, wherein, the predictor is M channel independent predictors, and is configured to make individual predictions on respective ones of the M first sequences respectively.
4. The system of any one of claims 1 to 3, wherein, each of the M refiners is configured to splice one of the M first sequences before or after a corresponding one of the M second sequences.
5. The system of any one of claims 1 to 3, wherein, the M refiners each comprise a layer of linear neural networks.
6. The system of any one of claims 1 to 3, wherein, training the system comprises training the predictor first, and training the refiner model after training the predictor while keeping parameters of the predictor unchanged.
7. The system of claim 6, wherein, the predictor is trained by minimizing mean squared error or absolute value error between the M second sequences and the true sequences, and wherein the refiner model is trained by minimizing mean squared error or absolute value error between the optimized M second sequences and the true sequences.
8. A method for training the system of any one of claims 1 to 5, comprising: training the predictor such that M second sequences output by the predictor approximate the true sequences; and training the refiner model after training the predictor while keeping parameters of the predictor unchanged, such that optimized M second sequences output by the refiner model approximate the true sequences.
9. An apparatus for training the system of any one of claims 1 to 5, comprising: first training means configured to train the predictor such that M second sequences output by the predictor approximate the true sequences; and second training means configured to train the refiner model after training the predictor while keeping parameters of the predictor unchanged, such that optimized M second sequences output by the refiner model approximate the true sequences.
10. A computer readable storage medium storing a program which, when executed by a processor, causes a computer to implement the method of claim 8.