Time sequence data prediction based on three-dimensional fully-connected fusion, and model training

Abstract

Disclosed in the present description are a time sequence data prediction method based on three-dimensional fully-connected fusion, and a model training method. One example of the method comprises: acquiring historical state data on the basis of original state data of a plurality of objects involved in a target task at a plurality of historical moments, and determining corresponding inherent features, time features and state features on the basis of the historical state data; then, performing data fusion of the inherent features, the time features and the state features by means of a target model, and performing data association on fused features obtained by fusion to obtain associated features; and then, predicting on the basis of the associated features by means of the target model to obtain prediction result data, and training the target model with an optimization objective of minimizing a deviation between the prediction result data and the original state data. Therefore, the trained target model can be used to predict state data to be predicted, and the target task is executed on the basis of a prediction result.

Description

Time series data prediction and model training based on 3D full-connection fusion Technical Field

[0001] This specification relates to the field of data analysis technology, and in particular to time series data prediction and model training based on three-dimensional fully connected fusion. Background Art

[0002] With the continuous development of science and technology, the application of data forecasting technology is becoming more and more widespread, ranging from daily weather forecasts to forecasts of market economic trends. This has led to an increasing emphasis on the accuracy and efficiency of data forecasting technology.

[0003] Currently, the primary method for data prediction involves leveraging various trained neural networks to predict future states based on known historical data, and then adjusting subsequent task execution based on the predicted results. While this approach can achieve the goal of predicting unknown states in specific scenarios, it generally suffers from low prediction accuracy and efficiency.

[0004] Summary of the Invention

[0005] This specification provides a time series data prediction method and model training method based on three-dimensional fully connected fusion, in order to relatively accurately predict data based on multivariate time series.

[0006] According to a first aspect of the present specification, a method for model training is provided, comprising: obtaining historical state data based on original state data of each object in a plurality of objects involved in a target task at each of a plurality of historical moments, wherein the historical state data is the same as the corresponding original state data, or is missing some data compared with the corresponding original state data; inputting the historical state data into a target model to be trained to generate inherent features, time features, and state features corresponding to the historical state data; performing data fusion on the inherent features, the time features, and the state features through the target model to obtain fused features corresponding to the historical state data; inputting the fused features corresponding to the historical state data into a data association layer of the target model to perform data association on the inherent features, the time features, and the state features in the fused features to generate associated features corresponding to the historical state data; performing data prediction based on the associated features corresponding to the historical state data through the target model to obtain predicted result data corresponding to the historical state data; and training the target model with the optimization goal of minimizing the deviation between the predicted result data corresponding to the historical state data and the original state data corresponding to the historical state data.

[0007] Optionally, the historical status data is input into the target model to be trained to generate inherent features corresponding to the historical status data, specifically including: inputting the historical status data into the target model to predict missing data in the historical status data through the target model, and completing the historical status data based on the prediction results; determining the inherent features corresponding to the historical status data based on the completed historical status data.

[0008] Optionally, the historical status data is input into the target model so as to predict the missing data in the historical status data through the target model, and the historical status data is completed according to the prediction result, specifically including: using preset instruction characters to mark the data bits of the missing data in the historical status data to obtain the historical status data after the mark is missing; inputting the historical status data after the mark is missing into the target model, so that the target model determines the data missing bits of the missing data according to the preset instruction characters in the historical status data, and predicts the data on the data missing bits to obtain the completed historical status data.

[0009] Optionally, the data association layer includes an inherent data association layer, a time data association layer, and a state data association layer. Accordingly, the fusion features corresponding to the historical state data are input into the data association layer of the target model to perform data association on the inherent features, the time features, and the state features in the fusion features to generate association features corresponding to the historical state data, specifically including: sending the fusion features to the inherent data association layer of the target model, so that the target model maps the fusion features based on the inherent features to obtain first mapping data; sending the first mapping data to the state data association layer of the target model, so that the target model maps the first mapping data based on the state features to obtain second mapping data; sending the second mapping data to the time data association layer of the target model, so that the target model maps the second mapping data based on the time features to obtain association features corresponding to the fusion features.

[0010] According to the second aspect of this specification, a time series data prediction method based on three-dimensional fully connected fusion is provided, which includes: obtaining state data to be predicted involved in a target task; inputting the state data to be predicted into a target model to obtain prediction result data for the state data to be predicted through the target model, wherein the target model is trained by the model training method as described above; and executing the target task according to the prediction result data for the state data to be predicted.

[0011] According to the third aspect of the present specification, a device for model training is provided, which includes: an acquisition module for acquiring historical state data based on the original state data of each object in multiple objects involved in the target task at each historical moment in multiple historical moments, wherein the historical state data is the same as the corresponding original state data, or is missing some data compared with the corresponding original state data; a feature generation module for inputting the historical state data into the target model to be trained to generate inherent features, time features and state features corresponding to the historical state data; a feature fusion module for fusing the inherent features, the time features and the state features through the target model to obtain the A fusion feature corresponding to the historical state data; a feature association module, used to input the fusion feature corresponding to the historical state data into the data association layer of the target model, so as to perform data association on the inherent feature, the time feature and the state feature in the fusion feature, and generate the association feature corresponding to the historical state data; a prediction module, used to perform data prediction based on the association feature corresponding to the historical state data through the target model, so as to obtain the prediction result data corresponding to the historical state data; a training module, used to train the target model with the optimization goal of minimizing the deviation between the prediction result data corresponding to the historical state data and the original state data corresponding to the historical state data.

[0012] Optionally, the feature generation module is specifically used to: input the historical status data into the target model to predict missing data in the historical status data through the target model, and complete the historical status data based on the prediction results; determine the inherent features corresponding to the historical status data based on the completed historical status data.

[0013] According to the fourth aspect of this specification, a device for time series data prediction based on three-dimensional fully connected fusion is provided, which includes: an acquisition module for acquiring state data to be predicted involved in a target task; a prediction module for inputting the state data to be predicted into a target model to obtain prediction result data for the state data to be predicted through the target model, wherein the target model is trained by the model training method as described above; and an execution module for executing the target task according to the prediction result data for the state data to be predicted.

[0014] According to a fifth aspect of this specification, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for model training and / or the method for time series data prediction based on three-dimensional fully connected fusion are implemented.

[0015] According to a sixth aspect of this specification, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the computer program implements the aforementioned model training method and / or the aforementioned method for predicting time series data based on three-dimensional fully connected fusion.

[0016] From the above content, it can be seen that the model training method and the time series data prediction method based on three-dimensional fully connected fusion provided in this specification, by utilizing the data association layer in the target model, associate the inherent features, time features and state features corresponding to the historical state data, so that data prediction can be performed based on the associated features that are closely related to each feature, making the prediction result data obtained based on the historical state data more accurate, effectively ensuring the smooth execution of the target task based on the prediction result data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] FIG1 is a flow chart of a model training method provided according to an embodiment of the present disclosure.

[0018] FIG2 is a schematic diagram of a method for completing historical status data with missing data according to an embodiment of the present specification.

[0019] FIG3 is a flow chart of a method for predicting time series data based on three-dimensional fully connected fusion according to an embodiment of this specification.

[0020] FIG4 is a schematic diagram of a model training device provided according to an embodiment of this specification.

[0021] FIG5 is a schematic diagram of a device for predicting time series data based on three-dimensional fully connected fusion according to an embodiment of this specification.

[0022] FIG6 is a schematic structural diagram of an electronic device provided according to an embodiment of this specification. DETAILED DESCRIPTION

[0023] The technical solutions provided by the embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0024] FIG1 is a flow chart of a model training method according to an embodiment of the present specification, which may specifically include the following steps.

[0025] In step S101, historical state data is obtained based on original state data of each of multiple objects involved in a target task at each of multiple historical moments, wherein the historical state data is the same as the corresponding original state data or lacks some data compared to the corresponding original state data.

[0026] In step S102 : the historical state data is input into a target model to be trained to generate inherent features, time features and state features corresponding to the historical state data.

[0027] With the continuous progress and innovation of the times, the demand for data prediction capabilities is gradually increasing in various fields involving data processing. However, currently, predicting the future based on historical data is mostly achieved through various pre-trained neural networks. This is generally plagued by low prediction efficiency and low accuracy, and there is a high probability of negatively impacting the subsequent execution of the target tasks based on the prediction results. Therefore, how to efficiently and accurately predict data based on historical data is crucial.

[0028] To this end, according to the embodiments of this specification, a model training method and a time series data prediction method based on three-dimensional full connection fusion are provided. The execution subject of the method can be a terminal device such as a desktop computer, a laptop computer, etc., or a server. In addition, the execution subject of the method can also be a subject in the form of software, such as a client installed in a terminal device. For the sake of convenience, the following only uses the terminal device as the execution subject to illustrate the model training method provided according to the embodiments of this specification and the time series data prediction method based on three-dimensional full connection fusion. Based on this, by applying the model training method provided according to the embodiments of this specification and the time series data prediction method based on three-dimensional full connection fusion, the terminal device can perform data prediction based on the acquired status data, obtain corresponding prediction result data, and then perform the corresponding target task based on the prediction result data.

[0029] The target tasks performed by the terminal device can be determined according to the actual scenario. For example, in the scenario of urban road traffic speed limit planning, the terminal device can predict the future vehicle speed of each traffic section based on the average vehicle speed collected from each traffic section in a certain time period from multiple traffic sections, so as to adjust the speed limit and other relevant regulations of each traffic section based on the prediction result data to reduce traffic congestion or traffic accidents. For another example, in the scenario of large-scale circuit voltage distribution, the terminal device can predict the future required working voltage of each power facility based on the working voltage of each power facility in different time periods obtained from multiple power facilities, and make timely adjustments and changes to the subsequent voltage distribution based on the prediction results to ensure the power safety of the entire circuit and energy saving and consumption reduction.

[0030] It can be seen that the method provided in the embodiments of this specification is mainly divided into two stages, namely the model training stage and the actual application stage. In the model training stage, the terminal device can obtain historical state data based on the original state data of each object in the multiple objects involved in the target task at each historical moment in multiple historical moments (wherein the historical state data is the same as the corresponding original state data or is missing some data compared with the corresponding original state data), and input the historical state data into the target model to be trained preset in the terminal device, so as to generate the inherent features, time features and state features corresponding to the historical state data through the target model.

[0031] Specifically, the terminal device can represent the acquired historical status data in the form of tensor data according to the two dimensions of time and object. For example, the tensor data shape corresponding to the specific historical status data can be expressed as the following formula (1): X = (B, L, N, Nd) (1).

[0032] Among them, "B" represents the number of batches divided equally according to the data length of the historical state data, "L" represents the length of the time series corresponding to the historical state data (hereinafter also referred to as the number of nodes), "N" represents the number of objects involved in the historical state data, and "Nd" represents the length of the feature dimension corresponding to the historical state data (that is, the number of associated features to be generated based on the historical state data and ultimately used for model training and / or data prediction, which can also be understood as the number of feature vectors formed by adding together the inherent features, time features and state features corresponding to the historical state data obtained based on the historical state data). It should be noted that the values ​​of the parameters representing the dimensions of tensor data can be dynamically adjusted according to actual needs, and this manual does not strictly limit them.

[0033] To illustrate the specific representation of historical status data in the form of tensor data, the following example is introduced in the field of traffic speed measurement. In particular, the specific historical status data in the form of tensor data in the field of traffic speed measurement can be referred to as follows:

[0034] Based on the historical state data in equation (2), the time series length L corresponding to the historical state data X is 4, that is, the number of time nodes L at which data is collected is 4, which can be named the first time node, the second time node, the third time node, and the fourth time node, respectively. The number of objects N involved in the historical state data X is 3, which can be named traffic speed sensor A, traffic speed sensor B, and traffic speed sensor C, respectively. These three traffic speed sensors are located on different test sections. The tensor form of the historical state data X can be expressed as: X = (B, 4, 3, 1) (3).

[0035] In the historical state data shown in formula (2), the specific value of the lowest dimension data is used to represent the average speed of passing vehicles collected by a certain object (traffic speed sensor) at the corresponding time node, in kilometers per hour. Specifically, taking the second row of data [[50.32], [79.56], [25.55]] in formula (2) as an example, this data row indicates that traffic speed sensor A collected an average speed of 50.32 kilometers per hour at the second time node (for example, 10:30 am), traffic speed sensor B collected an average speed of 79.56 kilometers per hour at the second time node, and traffic speed sensor C collected an average speed of 25.55 kilometers per hour at the second time node.

[0036] Before the target model generates the inherent features, time features and state features corresponding to the historical state data based on the historical state data, the terminal device will perform pre-processing such as data recognition on the historical state data in the form of tensor data, including judging whether there are null values ​​in each historical state data (that is, the data bits that should have data are missing for some reason). If the terminal device determines that there are missing data in the historical state data, the data bits of the missing data in the historical state data (hereinafter referred to as data missing bits) will be marked according to the preset instruction characters to obtain the historical state data after the mark is missing. For example, the preset instruction character can specifically be a set character such as the value "0", which is not specifically limited in this specification.

[0037] After marking the historical state data with missing data, the terminal device will also perform regularization processing on the historical state data in the form of tensor data to facilitate subsequent data processing. For example, the regularization processing can be specifically performed with reference to the following formula (4): Xregular = (X–mean(X)) / std(X) (4).

[0038] Where X represents the historical state data in the form of tensor data, mean(X) represents the mean of the historical state data, and std(X) represents the standard deviation of the historical state data. The regularization process is to subtract mean(X) from the historical state data X and then divide it by the standard deviation std(X) to obtain the regularized historical state data Xregular.

[0039] The terminal device then inputs the historical state data, which has been normalized and marked with missing data, into the target model. The target model then generates a zero-value mask tensor corresponding to the historical state data with missing data. The target model then predicts the original data corresponding to the missing data bits based on the zero-value mask tensor to obtain the completed historical state data without missing data. The following describes how to complete historical state data with missing data, using Figure 2.

[0040] FIG2 is a schematic diagram of a method for completing historical status data with missing data according to an embodiment of the present specification.

[0041] As shown in Figure 2, suppose that the tensor data X is a certain historical state data. For some reason, the data bits originally "6" and "8" are missing. The terminal device marks the data of the corresponding data bits as zero, that is, the historical state data after the missing mark in Figure 2. The historical state data after the missing mark is input into the target model to confirm the zero-value mask tensor corresponding to the historical state data after the missing mark through the target model. Specifically, as shown in Figure 2, the data bits with real data are represented by "false", and the data bits marked as missing data (that is, the data is "0") are represented by "true". Then, the target model is used to predict the original data corresponding to the missing data bits based on the zero-value mask tensor to complete the historical state data with missing data. For details, please refer to the following formulas (5) and (6): null_embdding=nn.Parameter(torch.zeros(1,1,N,1)) (5), X+=null_embedding.repeat(B,L,1,1)[X_mask] (6).

[0042] Among them, null_embdding represents a learnable variable used by the target model to predict missing data in the historical state data. During the model training process, the target model will automatically adjust the learnable variable null_embdding so that it becomes more and more similar to the real original data in the process of filling in the missing data. The specific replacement process can be shown in the above formula (6). According to the data distribution of "false" and "true" in the zero-value mask tensor [X_mask], the data bits in the historical state data X that need to be replaced by the learnable variable null_embdding are determined, that is, the data missing bits in the historical state data X, and the learnable variable null_embdding is replaced with the corresponding data missing bits. In simple terms, it is to complete the historical state data X by using the learnable variable null_embdding as the data corresponding to the data missing bits.

[0043] Because the terminal device has already regularized the historical state data after marking missing bits before using the target model to predict the data at the data-missing bits in the historical state data, the terminal device can directly use the completed historical state data as the inherent feature corresponding to the historical state data. It should be noted that if the terminal device does not find that there are missing bits in the historical state data, then the terminal device will directly regularize the historical state data without missing bits. The specific process is the same as the above-mentioned regularization process for the historical state data after marking missing bits, and input the regularized historical state data into the target model. In this case, the target model will use the regularized historical state data as the inherent feature corresponding to the historical state data without missing bits.

[0044] Next, the target model can determine the time features corresponding to the historical state data based on different time nodes under the time dimension. The time feature is mainly used to represent the historical moments corresponding to the historical state data, and can also be data in the form of tensors. The specific data representation of the time feature can be referred to the following formula (7): X_Stamp = torch.arange(T, dtype = torch.long) % 288 (7).

[0045] Among them, X_Stamp represents the time feature of the historical status data X; "288" means that 288 data recording points are determined with 5 minutes as a data recording point within the time range of 24 hours (0 o'clock to 24 o'clock); T represents each different time node under the time dimension. According to the above formula, the ordinal number of each different time node T under the time dimension of the historical status data X, which is a total of 288 data recording points within the time range of 24 hours (0 o'clock to 24 o'clock), is used as the time feature X_Stamp corresponding to the historical status data X. It should be noted that the number of data recording points and the time range mentioned above can be determined according to the actual needs of the actual application scenario, and this manual does not impose strict numerical limitations on them.

[0046] Then, the target model can determine the state features corresponding to the historical state data based on the different object information under the object dimension. The state features are mainly used to represent the state information of different objects involved in the historical state data, and can also be data in the form of tensors. The specific data representation of the state features can be referred to the following formula (8): node_embedding = nn.Parameter(torch.zeros(1,1,N,D)) (8).

[0047] Here, node_embedding represents a learnable tensor used by the target model to determine the state features corresponding to historical state data. During model training, the target model automatically adjusts the learnable tensor node_embedding to more accurately represent the state information of different objects within the object dimension, allowing for the interaction between objects in subsequent data prediction. D represents the size of the hidden layer in the target model and can be set and adjusted based on the actual needs of the application scenario.

[0048] The specific determination process and representation of the state features corresponding to the historical state data will be described using the example of the traffic speed measurement field. In particular, the specific data representation of the state features can be referred to the following formula (9): X node_embedding = [[2], [3], [1]] (9).

[0049] Among them, X node_embedding represents the state characteristics corresponding to each object (each traffic speed sensor) in the field of traffic speed measurement. For ease of understanding, it is assumed here that the detailed values ​​in the state characteristics only reflect the speed of the passing vehicles at the location of each traffic speed sensor. From the historical state data shown in formula (2) of the above example, it can be seen that the speed of the passing vehicles at the location of traffic speed sensor C is generally slow, while the speed of the passing vehicles at the location of traffic speed sensor B is generally fast. Therefore, [1] is used to represent the state characteristics of traffic speed sensor C, and [3] is used to represent the state characteristics of traffic speed sensor B. The above introduction process is only for ease of understanding. In actual application, the target model will consider the multi-faceted data of each object in the historical state data, thereby generating state characteristics that can better represent the characteristics of each object itself.

[0050] In step S103 : the inherent features, time features and state features are fused through the target model to obtain fused features corresponding to the historical state data.

[0051] In this specification, the terminal device can fuse the inherent features, time features, and state features corresponding to the historical state data through a preset target model. In particular, the target model can fuse multiple features in the form of tensor data into a fused feature in the form of tensor data.

[0052] Specifically, the terminal device can map the inherent features corresponding to the historical state data to a low-dimensional vector space through the target model. For example, the dimension of the inherent features corresponding to the historical state data can be reduced through the hidden layer preset in the target model to facilitate the subsequent data fusion process. For details, please refer to the following formulas (10) and (11): X1=XinherentW, (10) W∈R(1,D)

[0053] Wherein, Xinherent represents the inherent features corresponding to the historical state data X (for example, the historical state data after regularization as described above); W represents the weight matrix used to reduce the dimension of the inherent feature Xinherent, and its shape is (1, D); D represents the size of the hidden layer in the target model, and the hidden layer preset in the target model is mainly used to map high-dimensional tensor data to a low-dimensional vector space, thereby reducing the complexity and computational complexity of the target model; X1 represents the mapping result of the inherent feature Xinherent corresponding to the historical state data X in the low-dimensional vector space. It should be noted that the size of the hidden layer preset in the target model, that is, the specific value of D, can be determined according to the actual needs in the actual application scenario and is not limited in this specification.

[0054] Then, the terminal device can also map the time features corresponding to the historical state data to a low-dimensional vector space through the target model. In particular, this processing process can be similar to the above-mentioned processing process of the inherent features corresponding to the historical state data. For example, the dimension of the time features corresponding to the historical state data can be reduced through the hidden layer preset in the target model to facilitate the subsequent data fusion process. For details, please refer to the following formulas (12) and (13): time_embedding = nn.Embedding(288,D) (11). X2 = time_embedding(X_stamp) (12).

[0055] Here, X_stamp represents the time feature corresponding to the historical state data; time_embedding represents the time dictionary embedding layer in the target model, determined based on the number of data points and the time range; and X2 represents the mapping result of the time feature corresponding to the historical state data into the low-dimensional vector space. Based on the mapping data of the time feature X_stamp into the low-dimensional time dictionary embedding layer time_embedding, the mapping result X2 of the time feature into the low-dimensional vector space is determined.

[0056] After determining the inherent features corresponding to the historical state data and the mapping data of the time features in the low-dimensional vector space, the terminal device can fuse the determined mapping data with the state features to obtain the fusion features corresponding to the historical state data. For details, please refer to the following formula: Xfusion=X1+X2+node_embedding (13).

[0057] Here, X1 represents the mapping result of the inherent feature Xinherent corresponding to the historical state data X into the low-dimensional vector space, X2 represents the mapping result of the time feature X_stamp corresponding to the historical state data X into the low-dimensional vector space, node_embedding represents the state feature corresponding to the historical state data, and Xfusion represents the fusion feature corresponding to the historical state data. The main purpose of this process is to numerically add the various features of the historical state data X to obtain a fusion feature that simultaneously contains the inherent features, time features, and state features corresponding to the historical state data.

[0058] Due to the existence of the hidden layer of size D, the fusion feature in the form of tensor data corresponding to the historical state data can be expressed as shown in the following formula (14): Xfusion = (B, L, N, D) (14).

[0059] In this way, the high-dimensional historical state data is mapped into a low-dimensional space through the target model to obtain low-dimensional fusion features for subsequent data processing, which can effectively reduce the complexity and computational complexity of the target model and improve the generalization ability of the target model.

[0060] In step S104 , the fusion features corresponding to the historical state data are input into the data association layer of the target model to perform data association on the inherent features, time features and state features in the fusion features to generate association features corresponding to the historical state data.

[0061] In this specification, the terminal device can input the fusion features corresponding to the historical status data into the data association layer preset in the target model, and perform data association on the inherent features, time features and status features in the fusion features through the data association layer to obtain the association features corresponding to the historical status data.

[0062] The data association layer includes an inherent data association layer, a time data association layer, and a state data association layer. The terminal device first inputs the fusion features corresponding to the historical state data into the inherent data association layer of the target model, and then uses the inherent features in the fusion features as the standard through the inherent data association layer to map the fusion features to obtain the first mapping data. For details, please refer to the following formula (15):

[0063] Where X represents the fusion feature corresponding to the historical state data, W1 and W2 represent the two weight matrices in the mapping process for the inherent features, D represents the size of the preset hidden layer in the target model, and U represents the first mapping data after the mapping process for the inherent features based on the fusion feature. *,*,i It means that the inherent features in the fused features are used as the standard, the fused features X are mapped, and then the first mapping data U is obtained after the product processing of the weight matrix is ​​added.

[0064] Next, the terminal device may input the first mapping data into the state data association layer in the target model. The state data association layer may map the first mapping data based on the state feature in the fusion feature to obtain the second mapping data. For details, please refer to the following formula (16):

[0065] Where U represents the first mapping data corresponding to the fusion feature, W3 and W4 represent the two weight matrices in the mapping process for the state feature, N is the number of objects corresponding to the object dimension in the historical state data in the form of tensor data, and Y represents the second mapping data after the mapping process for the state feature based on the first mapping data. *,j,* It means that the first mapping data U is mapped based on the state feature, and then the second mapping data Y is obtained by adding the data after multiplying the data by the weight matrix.

[0066] It should be noted that before the state data association layer in the target model performs mapping processing on the state features, the tensor shape of the first mapping data needs to be dimensionally permuted to facilitate the mapping processing of the state features in the first mapping data by the state data association layer. For example, the object dimension in the first mapping data can be transposed to the last dimension of the first mapping data. For details, please refer to the following formula (17): X = X.permute(0,1,3,2) (17).

[0067] In the formula, the X to the right of the equal sign represents the tensor data that needs to be transposed, and permute(0,1,3,2) permutes the third and fourth dimensions of the tensor data. Specifically, combined with the aforementioned mapping of state features, permute(0,1,3,2) transposes the object dimension in the first mapping data to the last dimension of the first mapping data, allowing the state data association layer in the target model to directly map the state features in the first mapping data based on the data in the last dimension.

[0068] It should also be noted that after the state data association layer in the target model generates the second mapping data corresponding to the first mapping data, the dimensional order in the second mapping data needs to be permuted back to the original dimensional order in the first mapping data to ensure that the shape of the tensor data before the input state data association layer is consistent with the shape of the tensor data after the output. For example, the following formula (18) can be used for details: X = X.permute(0,1,3,2) (18).

[0069] Then, the terminal device can input the second mapping data into the time data association layer in the target model. The time data association layer uses the time feature in the fusion feature as the standard to perform mapping processing on the second mapping data to obtain the association feature. For details, please refer to the following formula:

[0070] Where Y represents the second mapping data corresponding to the fusion feature, W5 and W6 represent the two weight matrices in the mapping process for the time feature, L is the number of historical moments corresponding to the time dimension in the historical state data in the form of tensor data, and Z represents the associated features after the mapping process for the time feature based on the second mapping data. LayerNorm(Y) k,*,* It means that the second mapping data Y is mapped based on the time feature, and then the second mapping data Y is multiplied by the weight matrix and then added to obtain the associated feature Z.

[0071] It should also be noted that before the time data association layer in the target model performs mapping processing on the time features, the tensor shape of the second mapping data needs to be dimensionally permuted to facilitate the time data association layer to perform mapping processing on the time features in the second mapping data. For example, the time dimension in the second mapping data can be transposed to the last dimension of the second mapping data. For details, please refer to the following formula (20): X = X.permute(0,3,2,1) (20).

[0072] In the formula, the X to the right of the equal sign represents the tensor data that needs to be transposed, and permute(0,3,2,1) permutes the second and fourth dimensions of the tensor data. Specifically, combined with the aforementioned mapping of time features, permute(0,3,2,1) transposes the time dimension of the second mapped data to the last dimension of the second mapped data, allowing the time data association layer in the target model to directly map the time features in the second mapped data based on the data in the last dimension.

[0073] It should also be noted that after the time data association layer in the target model generates the associated features corresponding to the second mapping data, the dimension order in the associated features needs to be permuted back to the original dimension order in the second mapping data to ensure that the shape of the tensor data before the input time data association layer is consistent with the shape of the tensor data after the output. For details, refer to the following formula (21): X = X.permute(0,3,2,1) (21).

[0074] The purpose of the above process of data association for the fusion features corresponding to the historical state data is to make the various dimensions in the fusion features have more obvious correlations, thereby providing more reliable correlation features for subsequent data prediction through the target model.

[0075] In step S105 : data prediction is performed according to the associated features corresponding to the historical state data by using the target model to obtain prediction result data corresponding to the historical state data.

[0076] In step S106 , the target model is trained with minimizing the deviation between the prediction result data corresponding to the historical state data and the original state data corresponding to the historical state data as an optimization goal.

[0077] In this specification, the terminal device will use the fully connected layer in the target model to perform data prediction based on the associated features corresponding to each historical state data to obtain the predicted result data corresponding to the historical state data, and train the target model with the optimization goal of minimizing the deviation between the predicted result data and the corresponding original state data.

[0078] Specifically, the terminal device will input the associated features corresponding to the historical state data into the fully connected layer with data prediction function in the target model to perform data prediction through the fully connected layer to obtain the corresponding prediction result data. For details, please refer to the following formula (22): Y=FC(X.permute(0,2,1,3).view(B,N,L*D)) (22)

[0079] Here, X represents the associated features corresponding to the historical state data (i.e., the associated features Z mentioned in the data association process above); FC represents the fully connected layers in the target model; X.permute(0,2,1,3) swaps the order of the second and third dimensions of the tensor-formed associated features X; view(B,N,L*D) multiplies each historical moment in the historical state data with the data through the fully connected layers in the target model, representing the features of each historical moment as an L*D vector together with X.permute(0,2,1,3); and Y represents the predicted data corresponding to the associated features X. In summary, the above formula predicts the state data of each object at a future moment based on the associated features of the multi-dimensional combined historical state data.

[0080] Next, the terminal device may train the target model with the optimization goal of minimizing the deviation between the predicted result data and the corresponding original state data. Specifically, the following formula may be referred to:

[0081] Among them, Y represents the prediction result data generated by the target model based on the historical state data, Represents the original state data corresponding to the historical state data, and MSE represents the difference between the predicted result data Y and the original state data By adjusting the relevant parameters of each neural network layer in the target model with the goal of reducing the loss deviation MSE, the target model is optimized to train the target model.

[0082] It should be noted that this specification does not specifically limit the network architecture of the target model. For example, a neural network model such as a multilayer perceptron (MLP) may also be used. In fact, those skilled in the art can flexibly adjust the network architecture of the target model based on the actual needs and budget of the actual application scenario.

[0083] The methods described in this specification are divided into two main phases: the model training phase and the actual application phase. The model training phase described above is primarily used to obtain a trained target model with data prediction capabilities, so that in the actual application phase, the target model can be used to predict the state data to be predicted, obtain the corresponding prediction result data, and then perform the corresponding target task based on the prediction result data.

[0084] The following will illustrate the time series data prediction method based on three-dimensional fully connected fusion with reference to Figure 3.

[0085] FIG3 is a flow chart of a method for predicting time series data based on three-dimensional fully connected fusion provided in this specification, which includes the following steps:

[0086] In step S301: obtain the state data to be predicted involved in the target task.

[0087] With the continuous advancement of science and technology, the demand for and rigor of data prediction in various fields is increasing. However, existing data prediction methods often suffer from low efficiency and accuracy, which can severely impact the subsequent execution of tasks based on the predicted data. Therefore, how to accurately and efficiently predict historical data and successfully execute tasks based on the predicted results is an urgent problem to be solved.

[0088] To this end, this specification provides a time series data prediction method based on three-dimensional fully connected fusion. The execution subject adopted by the method can be a server or a terminal device such as a desktop computer or a laptop computer. In addition, the execution subject of the method can also be in the form of software, such as a client installed in a terminal device. For the sake of ease of explanation, the following only uses the server as the execution subject to illustrate the time series data prediction method based on three-dimensional fully connected fusion.

[0089] The specific target tasks to be performed can be determined according to the actual scenario. For example, in the field of weather forecasting, the target model can be used to predict weather data for a future time period based on past weather data, and then the weather forecast task can be performed based on the prediction results, including timely publicizing future weather changes and timely reminding citizens to take corresponding measures; for example, in the field of stock and securities trading, the target model can be used to predict the future direction of the stock market based on the historical change trends of each stock and the mutual influence between each stock, providing more comprehensive information for relevant personnel, so that the subsequent stock trading process can be adjusted. In this manual, the server can obtain the state data to be predicted and input the state data to be predicted into the target model in the form of tensor data.

[0090] The following is an example of a state data to be predicted in the form of tensor data in the field of weather forecasting. The specific example can be referred to the following formula (24):

[0091] Where X is the state data to be predicted, representing the temperature of three different regions at three different historical moments. From the data in formula (24), it can be seen that the time series length L corresponding to the state data to be predicted X is 3, that is, the number of time nodes for data collection is 3 (which can be respectively referred to as the first time node, the second time node, and the third time node); and the number of objects N involved in the state data to be predicted X is also 3, that is, the number of sensors for temperature collection is 3 (which can be respectively referred to as temperature sensor A, temperature sensor B, and temperature sensor C). Based on the above, the tensor shape of the state data to be predicted X shown in formula (24) can be expressed as: X = (B, 3, 3, 1) (25).

[0092] The specific value of the lowest-dimensional data is used to represent the temperature value collected by a temperature sensor at the corresponding historical moment, in degrees Celsius. Specifically, taking the second row of data [[23.4],[24.6],[23.1]] in formula (24) as an example, this data row indicates that the temperature collected by temperature sensor A at the second time node (e.g., 12 noon) is 23.4°C, the temperature collected by temperature sensor B at the second time node is 24.6°C, and the temperature collected by temperature sensor C at the second time node is 23.1°C.

[0093] In step S302, the state data to be predicted is input into a target model to obtain prediction result data for the state data to be predicted through the target model, wherein the target model is obtained by training using the above-mentioned model training method.

[0094] In step S303: executing the target task according to the prediction result data for the state data to be predicted.

[0095] In this specification, a server can input the state data to be predicted in the form of tensor data into a pre-trained target model. The target model then uses this data to associate the inherent features, state features, and time features corresponding to the state data to be predicted, obtaining associated features corresponding to the state data to be predicted. Based on these associated features, the server then performs data prediction on the state data to obtain prediction result data corresponding to the state data to be predicted. Subsequent target tasks can then be executed based on the prediction result data.

[0096] From the above content, it can be seen that the model training method and the time series data prediction method based on three-dimensional fully connected fusion provided in this specification, by utilizing the data association layer in the target model, associate the inherent features, time features and state features corresponding to the historical state data, so that data prediction can be performed based on the associated features that are closely related to each feature, and missing data can also be supplemented, so that the data used in the data prediction process is more comprehensive, and the results of the data prediction are more accurate.

[0097] The above are one or more implementation methods of this specification. Based on the same idea, this specification also provides a corresponding model training device, as shown in Figure 4.

[0098] FIG4 is a schematic diagram of a model training device provided in this specification, including an acquisition module 401 , a feature generation module 402 , a feature fusion module 403 , a feature association module 404 , a prediction module 405 and a training module 406 .

[0099] The acquisition module 401 is configured to acquire historical state data based on original state data of each of the multiple objects involved in the target task at each of the multiple historical moments, wherein the historical state data is the same as the corresponding original state data or is missing some data compared to the corresponding original state data.

[0100] The feature generation module 402 is used to input the historical state data into the target model to be trained to generate inherent features, time features and state features corresponding to the historical state data.

[0101] The feature fusion module 403 is used to perform data fusion on the inherent features, time features and state features through the target model to obtain fused features corresponding to the historical state data.

[0102] The feature association module 404 is used to input the fusion features corresponding to the historical state data into the data association layer of the target model to perform data association on the inherent features, time features and state features in the fusion features to generate association features corresponding to the historical state data.

[0103] The prediction module 405 is used to perform data prediction according to the associated features corresponding to the historical state data through the target model to obtain prediction result data corresponding to the historical state data.

[0104] The training module 406 is configured to train the target model with minimizing the deviation between the prediction result data corresponding to the historical state data and the original state data corresponding to the historical state data as an optimization goal.

[0105] Optionally, the feature generation module 402 is specifically used to: input the historical status data into the target model to predict the missing data in the historical status data through the target model; complete the historical status data based on the obtained prediction results; and determine the inherent features corresponding to each object based on the completed historical status data.

[0106] Optionally, the feature generation module 402 is further specifically used to: mark the data bits of missing data in the historical status data according to preset instruction characters to obtain the historical status data after the mark is missing; input the historical status data after the mark is missing into the target model, so that the target model determines the data missing bits of the missing data according to the preset instruction characters in the historical status data, and predicts the data on the data missing bits to obtain the completed historical status data.

[0107] Optionally, the data association layer includes an inherent data association layer, a time data association layer, and a state data association layer. Accordingly, the feature association module 404 is specifically used to: send the fused feature to the inherent data association layer of the target model, so that the target model maps the fused feature based on the inherent feature to obtain first mapping data; send the first mapping data to the state data association layer of the target model, so that the target model maps the first mapping data based on the state feature to obtain second mapping data; send the second mapping data to the time data association layer of the target model, so that the target model maps the second mapping data based on the time feature to obtain associated features corresponding to the fused feature.

[0108] Based on the same idea, this specification also provides a corresponding device for time series data prediction based on three-dimensional fully connected fusion, as shown in Figure 5.

[0109] FIG5 is a schematic diagram of a device for predicting time series data based on three-dimensional fully connected fusion provided in this specification, including an acquisition module 501 , a prediction module 502 and an execution module 503 .

[0110] The acquisition module 501 is used to acquire the state data to be predicted.

[0111] The prediction module 502 is used to input the state data to be predicted into a target model to obtain prediction result data for the state data to be predicted through the target model. The target model is trained by the above-mentioned model training method.

[0112] The execution module 503 is used to execute the target task according to the prediction result data for the state data to be predicted.

[0113] This specification also provides a computer-readable storage medium having a computer program stored thereon. The computer program can be used to execute the above-mentioned model training method and / or the time series data prediction method based on three-dimensional fully connected fusion.

[0114] This specification also provides a schematic structural diagram of an electronic device shown in Figure 6 that can be used to implement the above-mentioned model training method and / or the time series data prediction method based on three-dimensional fully connected fusion. As shown in Figure 6, at the hardware level, the electronic device 600 includes a processor 610, an internal bus 620, a network interface 630, a memory 640, and a non-volatile memory 650. Of course, the electronic device 600 may also include hardware required to support other services. The processor 610 reads the corresponding computer program from the non-volatile memory 650 and runs it in the memory 640 to implement the above-mentioned model training method or the above-mentioned time series data prediction method based on three-dimensional fully connected fusion.

[0115] In the 1990s, technological improvements could be clearly distinguished as either hardware improvements (for example, improvements to circuit structures like diodes, transistors, and switches) or software improvements (improvements to process flows). However, with the advancement of technology, many process flow improvements today can now be considered direct improvements to hardware circuit structures. Designers almost always create the corresponding hardware circuit structure by programming the improved process flow into the hardware circuit. Therefore, it cannot be said that a process flow improvement cannot be implemented using hardware modules. For example, a programmable logic device (PLD), such as a field programmable gate array (FPGA), is an integrated circuit whose logical function is determined by user programming. Designers can "integrate" a digital system on a PLD by programming it themselves, without having to hire a chip manufacturer to design and manufacture a dedicated integrated circuit chip. Moreover, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly done using "logic compiler" software. This is similar to the software compiler used when developing programs. Before compilation, the original code must also be written in a specific programming language, called a hardware description language (HDL). There is not just one HDL, but many, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art will also understand that by simply programming the method flow in one of these hardware description languages ​​and then programming it into an integrated circuit, a hardware circuit that implements the logic method flow can be easily obtained.

[0116] The controller can be implemented in any suitable manner. For example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also know that in addition to implementing the controller in a purely computer-readable program code format, the controller can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be considered as structures within the hardware component. Or even, the devices for implementing various functions can be considered as both software modules that implement the method and structures within the hardware component.

[0117] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0118] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0119] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0121] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0123] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-permanent storage in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium. Computer-readable media, including both permanent and non-permanent, removable and non-removable media, may be implemented using any method or technology for information storage. Information may be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change RAM (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, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device. According to the definition in this article, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0124] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0125] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Thus, this specification may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This specification may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0127] The various embodiments in this specification are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are generally similar to the method embodiments, so the description is relatively simple. For relevant parts, refer to the description of the method embodiments.

[0128] The foregoing is merely an example of the present invention and is not intended to limit the present invention. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A model training method, characterized in that: include: Acquire historical state data based on original state data of each object in a plurality of objects involved in the target task at each of a plurality of historical moments, wherein the historical state data is the same as the corresponding original state data or lacks some data compared with the corresponding original state data; Inputting the historical state data into a target model to be trained to generate inherent features, time features and state features corresponding to the historical state data; The inherent feature, the time feature and the state feature are subjected to data fusion through the target model to obtain a fusion feature corresponding to the historical state data; Inputting the fused features corresponding to the historical state data into the data association layer of the target model to perform data association on the inherent features, the time features, and the state features in the fused features to generate association features corresponding to the historical state data; Performing data prediction according to the associated features corresponding to the historical state data through the target model to obtain prediction result data corresponding to the historical state data; The target model is trained with the optimization goal of minimizing the deviation between the prediction result data corresponding to the historical state data and the original state data corresponding to the historical state data.

2. The method according to claim 1, characterized in that Inputting the historical state data into the target model to be trained to generate inherent features corresponding to the historical state data, specifically including: Inputting the historical state data into the target model, so as to predict the missing data in the historical state data through the target model, and completing the historical state data according to the prediction result; According to the completed historical status data, the inherent characteristics corresponding to the historical status data are determined.

3. The method according to claim 2, characterized in that Inputting the historical state data into the target model to predict missing data in the historical state data through the target model, and completing the historical state data according to the prediction result, specifically includes: Using a preset instruction character to mark the data bits of the missing data in the historical state data to obtain the historical state data after the mark is missing; The historical state data after the mark is missing is input into the target model, so that the target model determines the data missing bit of the missing data according to the preset instruction characters in the historical state data, and predicts the data on the data missing bit to obtain the completed historical state data.

4. The method according to claim 1, characterized in that The data association layer includes an inherent data association layer, a time data association layer and a state data association layer; Inputting the fusion features corresponding to the historical state data into the data association layer of the target model to perform data association on the inherent features, the time features, and the state features in the fusion features to generate association features corresponding to the historical state data, specifically including: Sending the fused features to the inherent data association layer of the target model, so that the target model performs mapping processing on the fused features based on the inherent features to obtain first mapping data; Sending the first mapping data to the state data association layer of the target model, so that the target model performs mapping processing on the first mapping data based on the state feature as a standard to obtain second mapping data; The second mapping data is sent to the time data association layer of the target model, so that the target model performs mapping processing on the second mapping data based on the time feature as a standard to obtain the association feature corresponding to the fusion feature.

5. A time series data prediction method based on three-dimensional full connection fusion, comprising: Obtain the state data to be predicted involved in the target task; Inputting the state data to be predicted into a target model to obtain prediction result data for the state data to be predicted through the target model, wherein the target model is trained by the method as described in any one of claims 1 to 4 above; The target task is executed according to the prediction result data for the state data to be predicted.

6. A model training device, comprising: An acquisition module, used to acquire historical state data based on original state data of each object in a plurality of objects involved in a target task at each of a plurality of historical moments, wherein the historical state data is the same as the corresponding original state data or lacks some data compared with the corresponding original state data; A feature generation module, used for inputting the historical state data into a target model to be trained to generate inherent features, time features and state features corresponding to the historical state data; A feature fusion module, used to fuse the inherent feature, the time feature and the state feature through the target model to obtain a fusion feature corresponding to the historical state data; A feature association module, used for inputting the fused features corresponding to the historical state data into the data association layer of the target model, so as to perform data association on the inherent features, the time features and the state features in the fused features, and generate association features corresponding to the historical state data; A prediction module, used to perform data prediction according to the associated features corresponding to the historical state data through the target model to obtain prediction result data corresponding to the historical state data; The training module is used to train the target model by taking minimizing the deviation between the prediction result data corresponding to the historical state data and the original state data corresponding to the historical state data as the optimization goal.

7. The device according to claim 6, characterized in that The feature generation module is specifically used for: Inputting the historical state data into the target model, so as to predict the missing data in the historical state data through the target model, and completing the historical state data according to the prediction result; According to the completed historical status data, the inherent characteristics corresponding to the historical status data are determined.

8. A device for predicting time series data based on three-dimensional fully connected fusion, comprising: An acquisition module is used to acquire the state data to be predicted involved in the target task; A prediction module, used for inputting the state data to be predicted into a target model to obtain prediction result data for the state data to be predicted through the target model, wherein the target model is trained by the model training method as described in any one of claims 1 to 4 above; An execution module is used to execute the target task according to the prediction result data for the state data to be predicted.

9. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the model training method described in any one of claims 1 to 4.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for predicting time series data based on three-dimensional fully connected fusion as described in claim 5 above is implemented.

11. An electronic device, comprising: Memory; processor; and A computer program stored in a memory and executable on a processor, Wherein, when the processor executes the program, the model training method described in any one of claims 1 to 4 is implemented.

12. An electronic device comprising: Memory; processor; and A computer program stored in a memory and executable on a processor, Wherein, when the processor executes the program, it implements the time series data prediction method based on three-dimensional full connection fusion described in claim 5 above.

Images