Implementation method and device of DMLP time sequence model, electronic equipment and storage medium
By introducing multi-layer perceptron and online adaptation layers into the DLinear algorithm, the problem of insufficient expression ability and adaptability is solved, and stronger prediction accuracy and adaptability is achieved, the fixed length limitation is broken, and the DMLP model is improved. Applicability.
Patent Information
- Application Number
- CN202411058467.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-29
AI Technical Summary
The existing DLinear algorithm has a simple network structure and a small number of parameters, which limits its expression ability, resulting in weak generalization ability and inability to adapt to data changes online. It requires frequent retraining, and the historical data length and prediction length are fixed, which limits its application. .
The multi-layer perceptron MLP is used to replace the linear structure in the DLinear network, add an online adaptation layer, use offline data training models, and fine-tune the actual data to achieve multi-step prediction and length interception to adapt to prediction tasks of different lengths.
It improves the expression ability and prediction accuracy of the model, breaks through the limitations of fixed historical data length and prediction length, has online fine-tuning capabilities, can better adapt to the time feature changes of the time series, and improves prediction performance and practicality.
Smart Images

Figure CN119047525A_ABST
Abstract
Description
Technical Field
[0001] This document relates to the field of computer technology, and in particular to a method, device, electronic device and storage medium for implementing a DMLP timing model. Background Art
[0002] Time Series Algorithms (TSA) are a class of algorithms specifically designed for analyzing and predicting time series data. Time series data is a series of data points recorded in chronological order. They can be observations of different time scales such as minutes, hours, days, months or years. Time series algorithms are widely used in many fields such as finance, meteorology, sales forecasting, resource management, etc. Common time series algorithms include: ARIMA (Autoregressive Integrated Moving Average Model), Exponential Smoothing, Seasonal Decomposition Time Series Forecasting (STL), XGBoost / LightGBM, Prophet, and some deep learning models. The choice of which time series algorithm depends on the nature of the specific problem, the characteristics of the data, and the prediction goals. In practical applications, it may be necessary to try multiple algorithms and determine which algorithm performs best on a specific data set through cross-validation.
[0003] Traditional time series algorithms include methods based on statistics, machine learning or deep learning, among which deep learning algorithms are more effective. They are usually based on CNN and RNN algorithms. These network structures cannot discover the long-term dependencies between time series data points, so the long-term prediction effect is poor. In addition, each sequence needs to be trained before prediction, resulting in low prediction efficiency. These have limited its practical application to a certain extent.
[0004] In recent years, the introduction of Transformer technology has led to great progress in the fields of NLP and large models, and many scholars have also applied this technology to time series tasks. The self-attention mechanism of Transformer allows it to discover regular patterns between time series data points and better utilize the long-term dependencies between data points. Therefore, many time series model algorithms based on this technology have achieved good prediction performance, far superior to traditional algorithms. Among them, the more outstanding ones include: iTransformer, TimesNet, NonStationaryFormer, etc. Compared with the technology using Transformer, algorithms based on simple linear transformations have also been found to have good performance, even surpassing most open source time series models. For example, the network structure of LTSFLinear is very simple, such as Figure 1 shown.
[0005] like Figure 2As shown in the figure, this super simple one-layer Linear network has achieved results that surpass many Transformer algorithms. Its network structure includes two types: DLinear and NLinear. The former first decomposes the sequence into trend terms and seasonal terms, and then uses Figure 1 The linear network is used to calculate, and then the two results are added to get the final prediction result. The latter is to first subtract the previous data point from each data point in the sequence to get the difference sequence, and then perform linear transformation on the difference sequence. Among them, Dlinear has better prediction performance.
[0006] In actual use, the Dlinear algorithm also has some problems. Its network structure is very simple and the number of parameters is very small, which is conducive to efficient deployment and high prediction efficiency, but it also severely limits its expressiveness and makes its generalization ability very weak. Moreover, the size of the linear transformation is strictly limited, equal to history length x prediction length. When the length of historical data or prediction length changes, it must be retrained, so it is basically one-dataset-one-model. Moreover, the training uses the historical data of the sequence, while the time series characteristics of the data in actual deployment will change over time. Dlinear does not have the ability of online training, so it cannot adapt to actual data changes, and even needs to be retrained and loaded with a new model every once in a while. Summary of the invention
[0007] The purpose of the present invention is to provide a method, device, electronic device and storage medium for implementing a DMLP timing model, aiming to solve the above-mentioned problems in the prior art.
[0008] The present invention provides a method for implementing a DMLP timing model, comprising:
[0009] The multi-layer perceptron MLP replaces the linear structure in the DLinear network to obtain the DMLP time series model containing MLP, where MLP is a 3-layer linear network model;
[0010] Adding an online adaptation layer to the output layer of the DMLP timing model, wherein the online adaptation layer is an identity mapping;
[0011] The DMLP time series model is trained using offline data to obtain a trained DMLP time series model, the trained DMLP time series model is deployed, and the DMLP time series model is fine-tuned and trained using the collected actual data.
[0012] The present invention provides a device for implementing a DMLP timing model, comprising:
[0013] A replacement module is used to replace the Linear structure in the DLinear network with a multi-layer perceptron MLP to obtain a DMLP time series model containing MLP, wherein the MLP is a 3-layer linear network model;
[0014] An adding module is used to add an online adaptation layer to the output layer of the DMLP timing model, wherein the online adaptation layer is an identity mapping;
[0015] The training module is used to train the DMLP time series model using offline data to obtain a trained DMLP time series model, deploy the trained DMLP time series model, and fine-tune the DMLP time series model using the collected actual data.
[0016] An embodiment of the present invention further provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the above-mentioned method for implementing the DMLP timing model when executed by the processor.
[0017] An embodiment of the present invention further provides a computer-readable storage medium, on which a program for implementing information transmission is stored, and when the program is executed by a processor, the steps of the method for implementing the above-mentioned DMLP timing model are implemented.
[0018] By adopting the embodiment of the present invention, the linear structure model prediction algorithm model DMLP is proposed, which has good expression ability, higher prediction accuracy, stronger generalization ability, breaks through the limitations of fixed historical data length and prediction length, and has the ability of online fine-tuning, which can better adapt to the changes in the time characteristics of the time series, and therefore has excellent prediction performance and good practicality. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0020] Figure 1 It is a schematic diagram of the network structure of LTSFLinear in the prior art;
[0021] Figure 2 It is a schematic diagram of the effect of a layer of Linear network algorithm in the prior art;
[0022] Figure 3is a flow chart of a method for implementing a DMLP timing model according to an embodiment of the present invention;
[0023] Figure 4 is a schematic diagram of a DMLP model according to an embodiment of the present invention;
[0024] Figure 5 is a schematic diagram of a device for implementing a DMLP timing model according to an embodiment of the present invention;
[0025] Figure 6 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0026] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.
[0027] Method Embodiment
[0028] According to an embodiment of the present invention, a method for implementing a DMLP timing model is provided. Figure 3 is a flow chart of a method for implementing a DMLP timing model according to an embodiment of the present invention. Figure 3 As shown, the implementation method of the DMLP timing model according to the embodiment of the present invention specifically includes:
[0029] Step S301, replacing the Linear structure in the DLinear network with a multi-layer perceptron MLP to obtain a DMLP time series model including the MLP, wherein the MLP is a 3-layer linear network model;
[0030] Step S302, adding an online adaptation layer to the output layer of the DMLP timing model, wherein the online adaptation layer is an identity mapping;
[0031] Step S303: Use offline data to train the DMLP time series model to obtain a trained DMLP time series model, deploy the trained DMLP time series model, and fine-tune the DMLP time series model using the collected actual data. The fine-tuning training specifically includes:
[0032] The trained DMLP timing model is deployed, and according to the actual data collected, other parameters of the DMLP timing model are fixed, and only the online adaptation layer of the DMLP timing model is trained to complete the fine-tuning training of the DMLP timing model.
[0033] When the historical data length and prediction length of the prediction task do not match the input and output length of the DMLP time series model, the prediction results or historical data are processed by multi-step prediction and / or length truncation.
[0034] Specifically include:
[0035] When the prediction length of the prediction task is greater than the preset length of the algorithm, the multi-step prediction method is used to process the prediction results;
[0036] When the prediction length of the prediction task is less than or equal to the preset length of the algorithm, the data of the required length is directly intercepted from the prediction result from front to back;
[0037] When the historical data length of the prediction task is greater than or equal to the historical data length of the algorithm, the data of the required length is directly intercepted from the historical data from back to front;
[0038] When the length of historical data of the prediction task is less than the length of historical data of the algorithm, the mean of the historical data is calculated, the mean is used to make up the length of the historical data, and it is added in front of the original historical data.
[0039] In summary, the embodiment of the present invention modifies and optimizes the network structure, prediction method, online fine-tuning and many other aspects of the DLinear algorithm to solve these problems in its use, and the new algorithm model obtained is called DMlp. The algorithm uses MLP (linear perceptron) instead of Linear structure, thereby greatly enhancing the representation ability of the model; for the trained model, an identity mapping is added to the output side so that the output of the model remains unchanged after passing through this mapping layer, which will serve as the model online adaptation layer (Online Reconciliation Layer); during the deployment process of the model, actual data is collected, cleaned and the model is fine-tuned and trained, and other parameters are fixed during training, and only the online adaptation layer is trained; further, when the historical data length and prediction length of the prediction task do not match the input and output length of the DMLP model, this limitation is broken through by multi-step prediction and length interception. Compared with the Dlinear algorithm, the DMlp of the embodiment of the present invention has good expression ability, higher prediction accuracy, stronger generalization ability, breaks through the limitations of fixed historical data length and prediction length, and has the ability of online fine-tuning, which can better adapt to the changes in the time characteristics of the time series, so it has excellent prediction performance and good practicality.
[0040] The above technical solution of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings.
[0041] The DMlp algorithm model uses MLP (multi-layer linear perceptron) instead of Linear structure, which greatly enhances the representation ability of the model.
[0042] The network structure of DLinear is very simple. Figure 1 As shown, its network structure is defined as:
[0043] self.Linear_Seasonal=nn.Linear(self.seq_len,self.pred_len)
[0044] self.Linear_Trend=nn.Linear(self.seq_len,self.pred_len)
[0045] Specifically, the network structure of DLinear consists of seasonal terms and trend terms, which are simple 2-layer linear networks like the first figure, where L = self.seq_len for history L timestamp and T = self.pred_len for future T timestamp. The nodes in these two layers are directly connected.
[0046] Because the network structure is very simple, it is very easy to overfit when the amount of data is small or the model is slightly complex.
[0047] MLP is the abbreviation of Multilayer Perceptron, which is a feedforward artificial neural network. MLP contains at least three layers of nodes: an input layer, an output layer, and at least one hidden layer. Except for the input node, each node is a neuron with a nonlinear activation function.
[0048] The DMLP model defined in the embodiment of the present invention is a three-layer network structure, and its network structure is:
[0049] self.Linear_Seasonal=nn.Sequential(nn.Linear(self.seq_len,self.d_model),nn.Linear(self.d_model,self.pred_len),
[0050] self.Linear_Trend=nn.Sequential(nn.Linear(self.seq_len,self.d_model),nn.Linear(self.d_model,self.pred_len)
[0051] Specifically, DMlp also consists of Seasonal and Trend, but the difference is that the 2-layer linear network is replaced by a 3-layer linear network. The 3-layer linear network consists of self.seq_len nodes in the first layer, self.d_model nodes in the middle layer, and self.pred_len nodes in the third layer. The nodes in the first and middle layers are connected in pairs, and the nodes in the middle and third layers are connected in pairs.
[0052] Among them, d_model is the size of the middle hidden layer, usually set d_model = 2*seq_len. Note that no nonlinear activation functions such as ReLU (Rectified Linear Unit), sigmoid or tanh, and regularization techniques such as Dropout are used here. Therefore, the MLP defined here is a 3-layer linear network model.
[0053] Compared with the Linear model, the MLP adds more weight parameters, so it has better expressive power and is less prone to overfitting.
[0054] For the trained DMLP model, an identity mapping is added on the output side so that the output of the model remains unchanged after passing through this mapping layer. It will serve as the model's online adaptation layer (Online Reconciliation Layer).
[0055] The trained DMLP model adds an identity mapping on the output side, and the output of the model remains unchanged after passing through this mapping layer. It will serve as the model's online adaptation layer (Online Reconciliation Layer), enabling the model to have online training / fine-tuning capabilities.
[0056] Self.Online_Recon_Layer_Seasonal=nn.Linear(self.pred_len,self.pred_len)
[0057] Self.Online_Recon_Layer_Trend=nn.Linear(self.pred_len,self.pred_len)
[0058] The size of this linear network is self.pred_len x self.pred_len. Specifically, the online layer is a simple two-layer linear network. Like seasonal or trend of Dlinear, its characteristic is that the number of nodes in the first layer and the second layer is the same, both equal to self.pred_len. In this way, the nodes in the two-layer network are connected in pairs, with a total of pred_len*pred_len lines, corresponding to pred_len*pred_len weight parameters, so here it is called the size (that is, the number of weights). In order to make it an identity mapping, we initialize it:
[0059] # Initialize weights to the identity matrix and bias to the zero vector
[0060] nn.init.eye_(Self.Online_Recon_Layer_Seasonal.weight)
[0061] nn.init.eye_(Self.Online_Recon_Layer_Trend.weight)
[0062] nn.init.Zeros_(Self.Online_Recon_Layer_Seasonal.bias)
[0063] nn.init.zeros_(Self.Online_Recon_Layer_Trend.bias)
[0064] Add the DMLP model of online_recon_layer as follows Figure 4 shown.
[0065] During the model deployment process, actual data is collected, cleaned and the model is fine-tuned. During training, other parameters are fixed and only the online adaptation layer is trained.
[0066] Usually, time series algorithm models are trained using offline data. After the trained model is deployed, it is usually not trained online. This causes the model to be unable to respond to changes in actual data over time, that is, it has no ability to adapt to environmental changes and cannot be continuously optimized.
[0067] If you want the model to be able to respond to environmental changes, you need to train the model online, which is divided into full parameter training and fine-tuning training. The former is to train all parameters in the model, which is usually more expensive; the latter is to train only a part of the parameters, such as LoRA, LadderSide, etc.
[0068] For our DMLP model, we choose to fix other parameters and only adjust the parameters of the online adaptation layer. This measure does not affect the parameters of the offline trained MLP model. Only the online adaptation layer, which is initially an identity mapping, is adjusted. The hierarchy is clear and the operation is simple. The number of online parameters is usually much smaller than that of MLP. If the adjustment efficiency is not good, it is also easy to restore to the initial state.
[0069] When the historical data length and prediction length of the prediction task do not match the input and output length of the DMLP model, this limitation can be overcome by multi-step prediction and length truncation.
[0070] The DLinear algorithm model requires fixed historical data length and prediction length. When the prediction task changes and these two length parameters change, the model needs to be retrained. This limitation is undoubtedly very large and is not conducive to the actual application of the algorithm model. In order to break through this limitation, in the application of the DMlp algorithm model, multi-step prediction and length truncation are used to break through this limitation:
[0071] When the prediction length > self.pred_len (the prediction length of the algorithm), self.seq_len represents the length of the algorithm input sequence, the multi-step prediction method is used:
[0072] Predict = []
[0073] While pred_len>0:
[0074] Predict_Data=DMLP(History_Data)
[0075] MinLen=Min(len(PredictData),self.pred_len)
[0076] Predict=Predict+PredcitData[MinLen:]
[0077] History_Data=History_Data+Predict_Data
[0078] History_Data=History_Data[-self.history_len:]
[0079] pred_len=pred_len-MinLen
[0080] return Predict
[0081] Multi-step prediction example:
[0082] It is required to predict 110 data, but the self.pred_len of the algorithm model is 50, that is, each prediction can only predict the next 50 future values; thus, 3 predictions are needed:
[0083] First, predict 50 future values;
[0084] Second, take the previous 50 values as part of the historical data and then predict the next 50 values; (in this way, 100 future values are obtained)
[0085] Third, similarly, take these 100 values as part of the historical data and then predict the next 50 values.
[0086] In this way, a total of 150 future values are predicted, and the first 110 of these values are taken.
[0087] For this set of operations, first perform multi-step prediction. If it exceeds the required length, then perform a truncation operation once.
[0088] When the prediction length < self.pred_len, simply directly truncate the required length from the front to the back in the prediction result
[0089] When the length of the historical data > self.history_len (the length of the historical data required by the algorithm), simply directly intercept the data of the required length from the back to the front in the historical data.
[0090] For example, the algorithm requires: use 300 historical data to predict 50 future values. self.history_len = 300, self.pred_len = 50.
[0091] When the length of the historical data < self.history_len, calculate the mean of these historical data, supplement the length of the historical data with the mean, and add it in front of the original historical data
[0092] In this way, in principle, for historical data of any length and prediction length of any length, the DMlp algorithm model can effectively perform predictions.
[0093] The algorithm model of the embodiment of the present invention has good expressive ability, stronger prediction accuracy, stronger generalization ability, breaks through the limitations of fixed historical data length and prediction length, and has the ability of online fine-tuning, and can better adapt to the time feature changes of the time series. Therefore, it has excellent prediction performance and good practicability.
[0094] Device Embodiment 1
[0095] According to an embodiment of the present invention, a device for implementing a DMLP timing model is provided. Figure 5 Schematic diagram of a device for implementing a DMLP timing model according to an embodiment of the present invention. Figure 5 As shown, the implementation device of the DMLP timing model according to the embodiment of the present invention specifically includes:
[0096] A replacement module 50 is used to replace the Linear structure in the DLinear network with a multi-layer perceptron MLP to obtain a DMLP time series model including the MLP, wherein the MLP is a 3-layer linear network model;
[0097] An adding module 52 is used to add an online adaptation layer to the output layer of the DMLP timing model, wherein the online adaptation layer is an identity mapping;
[0098] The training module 54 is used to train the DMLP time series model using offline data to obtain a trained DMLP time series model, deploy the trained DMLP time series model, and fine-tune the DMLP time series model using the collected actual data. The training module is specifically used for:
[0099] The trained DMLP timing model is deployed, and according to the actual data collected, other parameters of the DMLP timing model are fixed, and only the online adaptation layer of the DMLP timing model is trained to complete the fine-tuning training of the DMLP timing model.
[0100] The device further comprises:
[0101] The data processing module is used to process the prediction results or historical data by multi-step prediction and / or length truncation when the historical data length and prediction length of the prediction task do not match the input and output length of the DMLP time series model. The data processing module is specifically used to:
[0102] When the prediction length of the prediction task is greater than the preset length of the algorithm, the multi-step prediction method is used to process the prediction results;
[0103] When the prediction length of the prediction task is less than or equal to the preset length of the algorithm, the data of the required length is directly intercepted from the prediction result from front to back;
[0104] When the historical data length of the prediction task is greater than or equal to the historical data length of the algorithm, the data of the required length is directly intercepted from the historical data from back to front;
[0105] When the length of historical data of the prediction task is less than the length of historical data of the algorithm, the mean of the historical data is calculated, the mean is used to make up the length of the historical data, and it is added in front of the original historical data.
[0106] The embodiment of the present invention is a device embodiment corresponding to the above-mentioned method embodiment. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.
[0107] Device Example 2
[0108] An embodiment of the present invention provides an electronic device, such as Figure 6 As shown, it includes: a memory 60, a processor 62, and a computer program stored in the memory 60 and executable on the processor 62. When the computer program is executed by the processor 62, the steps described in the method embodiment are implemented.
[0109] Device Example 3
[0110] An embodiment of the present invention provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by the processor 62, the steps described in the method embodiment are implemented.
[0111] The computer-readable storage medium in this embodiment includes, but is not limited to, ROM, RAM, magnetic disk or optical disk, etc.
[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for implementing a DMLP timing model, characterized in that: include: The multi-layer perceptron (MLP) replaces the linear structure in the DLinear network to obtain a DMLP time series model containing MLP, where MLP is a 3-layer linear network model. Adding an online adaptation layer to the output layer of the DMLP time series model, wherein the online adaptation layer is an identity mapping; The DMLP time series model is trained using offline data to obtain a trained DMLP time series model, the trained DMLP time series model is deployed, and the DMLP time series model is fine-tuned and trained using the collected actual data.
2. The method according to claim 1, characterized in that The method further comprises: When the historical data length and prediction length of the prediction task do not match the input and output length of the DMLP time series model, the prediction results or historical data are processed through multi-step prediction and / or length truncation.
3. The method according to claim 2, characterized in that The processing of forecast results or historical data by multi-step forecasting and / or length truncation specifically includes: When the prediction length of the prediction task is greater than the preset length of the algorithm, the multi-step prediction method is used to process the prediction results; When the prediction length of the prediction task is less than or equal to the preset length of the algorithm, the data of the required length is directly intercepted from the prediction result from the front to the back; When the historical data length of the prediction task is greater than or equal to the historical data length of the algorithm, the data of the required length is directly intercepted from the historical data from back to front; When the historical data length of the prediction task is less than the historical data length of the algorithm, the mean of the historical data is calculated, the mean is used to make up the length of the historical data, and is added to the front of the original historical data.
4. The method according to claim 1, wherein Deploying the trained DMLP time series model and fine-tuning the DMLP time series model using the collected actual data specifically includes: The trained DMLP time series model is deployed. According to the actual data collected, other parameters of the DMLP time series model are fixed, and only the online adaptation layer of the DMLP time series model is trained to complete the fine-tuning training of the DMLP time series model.
5. A device for implementing a DMLP timing model, characterized in that: include: The replacement module is used to replace the linear structure in the DLinear network with a multi-layer perceptron (MLP) to obtain a DMLP time series model containing the MLP, where the MLP is a three-layer linear network model. An adding module, configured to add an online adaptation layer to the output layer of the DMLP time series model, wherein the online adaptation layer is an identity mapping; The training module is used to train the DMLP time series model using offline data to obtain a trained DMLP time series model, deploy the trained DMLP time series model, and fine-tune the DMLP time series model using the collected actual data.
6. The device according to claim 5, characterized in that The device further comprises: The data processing module is used to process the prediction results or historical data through multi-step prediction and / or length truncation when the historical data length and prediction length of the prediction task do not match the input and output length of the DMLP time series model.
7. The device according to claim 6, characterized in that The data processing module is specifically used for: When the prediction length of the prediction task is greater than the preset length of the algorithm, the multi-step prediction method is used to process the prediction results; When the prediction length of the prediction task is less than or equal to the preset length of the algorithm, the data of the required length is directly intercepted from the prediction result from the front to the back; When the historical data length of the prediction task is greater than or equal to the historical data length of the algorithm, the data of the required length is directly intercepted from the historical data from back to front; When the historical data length of the prediction task is less than the historical data length of the algorithm, the mean of the historical data is calculated, the mean is used to make up the length of the historical data, and is added to the front of the original historical data.
8. The device according to claim 5, characterized in that The training module is specifically used for: The trained DMLP time series model is deployed. According to the actual data collected, other parameters of the DMLP time series model are fixed, and only the online adaptation layer of the DMLP time series model is trained to complete the fine-tuning training of the DMLP time series model.
9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the method for implementing the DMLP timing model according to any one of claims 1 to 4 are implemented.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores an information transmission implementation program, and when the program is executed by a processor, the steps of the implementation method of the DMLP timing model according to any one of claims 1 to 4 are implemented.
