Long sequence time series forecasting apparatus based on a neural network having a CNN-MLP hybrid structure and method thereof

Abstract

The present invention relates to a long-term time series prediction device based on a neural network. The time series prediction device predicts time series data using a long-term time series prediction model based on a CNN-MLP hybrid model. The long-term time series prediction model comprises: a data embedding unit that patches-embeds input time series data; a CNN-MLP hybrid mixer unit that divides the embedded data into first and second data, performs a linear transformation on the first data and a convolution operation on the second data, and combines and fuses the linearly transformed first data and the convolutionally performed second data; and a head prediction unit that transforms the output of the CNN-MLP hybrid mixer unit into one dimension, performs a linear transformation, and predicts and outputs time series data having a target time series length (T).

Description

Technology Field

[0001] The present invention relates to a long-term time series prediction system and method based on a neural network, and more specifically, to a long-term time series prediction system, apparatus, and method configured to reduce the size of the model and enhance the processing capability for local information by utilizing a long-term time series prediction model modeled by training a neural network of a CNN-MLP hybrid structure. Background Technology

[0002] Time series data refers to a set of sequentially defined data collected over a specific period of time. These data are characterized by the fact that they have a temporal order and that consecutive observations are correlated with one another. The field of time series forecasting is utilized in a wide range of areas and has long been an actively researched field. Examples include financial market forecasting, weather forecasting, sensor data forecasting, energy consumption forecasting, and health data forecasting, and research related to these areas has been steadily conducted for a long time.

[0003] Time series forecasting refers to the study of predicting a time series at a future point in time based on a specific time series. Input time series is a time series of length L, and i represents a channel in a multivariate time series between 1 and C. The multivariate time series of length L is It is represented as. Time series forecasting is a time series of length T here. The goal is to predict. Generally, the predicted value is It is represented as, and this is the actual value Calculate the loss value using the loss function and update the model.

[0004] There are various deep learning models used for time series forecasting, mainly RNN, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), TCN, and Transformer models.

[0005] A recurrent neural network (RNN) is a neural network designed for sequential data. Figure 1 is a structural diagram illustrating an RNN neural network. As shown in Figure 1, the RNN has a structure that remembers the output of previous information and combines it with the next input to generate an output value. Due to this characteristic, it is frequently used in a Sequence-to-Sequence structure when performing time series forecasting. However, RNNs suffer from a lack of long-term memory, where previous information is lost as the data length increases. Due to this problem, the LSTM model was developed.

[0006] Long short-term memory (LSTM) is a neural network developed to compensate for the lack of long-term memory, a disadvantage of RNNs. Figure 2 is a structural diagram illustrating an LSTM neural network. Referring to Figure 2, the LSTM neural network manages memory using three gates consisting of input, output, and forget. Through this, it became possible to forget or remember past information. The LSTM showed a steep increase in the Mean Squared Error (MSE) score as the length to be predicted increased. However, the Transformer, a model with strengths in processing global information in data, emerged, and research using it was conducted.

[0007] A Temporal Convolutional Network (TCN) predicts time series using convolutional neural networks. Figure 3 is a structural diagram illustrating a TCN neural network. Referring to Figure 3, a single TCN layer consists of multiple dilated convolutional layers arranged hierarchically. Output at a single time point It takes only the time point prior to that point as input.

[0008] Figure 4 is a structural diagram illustrating a Transformer neural network. Referring to Figure 4, the Transformer model consists of an Encoder layer composed of a Multi-head Attention layer and a Feed Forward Network, and a Decoder in which Masked Multi-head Attention is added to the structure of the Encoder. Although the aforementioned Transformer model was initially proposed in natural language processing research, studies using it in image and time series fields have recently emerged. Time series forecasting research using the Transformer model has shown good performance compared to existing models. However, due to the characteristics of the model, using a pure Transformer model is difficult. Computational problems arise due to the complexity, leading to the development of various methods to address this. Representative examples include Reformer and LogTrans, which can reduce the time complexity of Transformer models.

[0009] However, due to the large number of parameters, Transformer models face the problem of requiring significant GPU memory when processing multivariate time-series data with many attributes, such as traffic data. This can cause issues when using the model in field applications where time-series data from various sources must be predicted simultaneously, rather than relying on a single dataset. Prior art literature

[0010] Korean Registered Patent Publication No. 10-2609569, Korean Published Patent Publication No. 10-2022-0150618, Korean Registered Patent Publication No. 10-2546334, Korean Registered Patent Publication No. 10-2480150 The problem to be solved

[0011] To solve the aforementioned problems, the present invention aims to provide a long-term time series prediction system and a method using a long-term time series prediction model based on a neural network having a CNN-MLP hybrid structure. means of solving the problem

[0012] A long-term time series prediction system based on a neural network according to the present invention for achieving the aforementioned technical problem comprises: a time series prediction device that predicts time series data using a long-term time series prediction model based on a CNN-MLP hybrid model; wherein the long-term time series prediction model comprises: a data embedding unit that patch-embeds input time series data; a CNN-MLP hybrid mixer unit that divides the embedded data into first and second data, performs a linear transformation on the first data and a convolution operation on the second data, and combines and fuses the linearly transformed first data and the convolutionally performed second data; and a head prediction unit that transforms the output of the CNN-MLP hybrid mixer unit into one dimension, performs a linear transformation, and predicts and outputs time series data having a target time series length (T).

[0013] In a long-term time series forecasting system based on a CNN-MLP hybrid model according to the aforementioned features, the data embedding unit comprises: a normalization module that converts the time series data into univariate time series data and normalizes it; a patch module that divides the normalized time series data into patches having a preset length; and a first linear layer that projects each patch onto a preset latent dimension using a first weight; and preferably, the input time series data is patch-embedded and output.

[0014] In a long-term time series forecasting system based on a CNN-MLP hybrid model according to the aforementioned features, the CNN-MLP hybrid mixer preferably comprises: a data splitting module that normalizes the patch-embedded data and then splits it into first data and second data in the latent dimension; a second linear layer that linearly transforms the first data using second weights; a convolution module that performs a convolution operation on the second data; a combination module that connects the first data linearly transformed by the second linear layer and the second data that has undergone a convolution operation by the convolution module; a convolution mixing module that combines information of the data connected by the combination module using depth-wise convolution; a Gated Attention module that linearly transforms the data mixed by the convolution mixing module and then applies a softmax function to generate output data; and a Feed Forward network that mixes information along the latent dimension axis for the output data of the Gated Attention module.

[0015] In a long-term time series forecasting system based on a CNN-MLP hybrid model according to the aforementioned features, the convolution module is preferably configured to learn information between adjacent patches by comprising: a third linear layer that linearly transforms data using a third weight; a depth-wise convolution layer that performs depth-wise convolution operations on the data linearly transformed by the third linear layer; and an activation layer that activates the data convolutionally performed using an activation function.

[0016] A long-term time series forecasting system based on a CNN-MLP hybrid model according to the aforementioned features further comprises a model training device that trains a long-term time series forecasting model based on a CNN-MLP hybrid model using training data including time series data and target time series data.

[0017] It is preferable that the above model training device denormalizes the data predicted from the time series data of the training data by a long-term time series prediction model, calculates the loss score of the denormalized predicted data and the actual target time series data using a loss function, and trains the long-term time series prediction model through backpropagation.

[0018] A long-term time series prediction method in a neural network-based long-term time series prediction system according to the present invention comprises: (a) a step of patch embedding input time series data; (b) a step of dividing the patch-embedding data into first and second data, then linearly transforming the first data and performing a convolution operation on the second data, and combining and fusing the linearly transformed data and the convolutionally performed data; and (c) a step of transforming the output of the CNN-MLP hybrid mixer into one dimension and then linearly transforming it to predict and output time series data having a target time series length (T).

[0019] In a long-term time series forecasting method based on a CNN-MLP hybrid model according to the aforementioned features, the step (a) comprises: (a1) a step of converting the time series data into univariate time series data and normalizing it; (a2) a step of dividing the normalized time series data into patches having a preset length; and (a3) ​​a step of projecting each patch onto a preset latent dimension using a first weight; thereby preferably outputting the input time series data by patch embedding.

[0020] In a long-term time series forecasting method based on a CNN-MLP hybrid model according to the aforementioned features, it is preferable that step (b) comprises: (b1) a step of normalizing the patch-embedded data and then splitting it into first data and second data in the latent dimension; (b2) a step of linearly transforming the first data using second weights; (b3) a step of performing a convolution operation on the second data; (b4) a step of connecting the linearly transformed first data and the convolutionally generated second data; (b5) a step of combining information of the data connected by the combination module using depth-wise convolution; (b6) a step of linearly transforming the data mixed by step (b5) and then applying a softmax function to generate output data; and (b7) a step of mixing information along the latent dimension axis for the output data of step (b6) through a feed-forward network.

[0021] In a long-term time series forecasting method based on a CNN-MLP hybrid model according to the aforementioned features, the step (b3) is preferably configured to learn information between adjacent patches by comprising: (b3-1) a step of linearly transforming data using a third weight; (b3-2) a step of performing a depth-wise convolution operation on the linearly transformed data; and (b3-3) a step of activating the convolutionally performed data using an activation function.

[0022] A long-term time series forecasting method based on a CNN-MLP hybrid model according to the aforementioned features further comprises: (d) a step of training a long-term time series forecasting model based on a CNN-MLP hybrid model using training data including time series data and target time series data; and

[0023] In the above step (d), it is preferable to denormalize the data predicted from the time series data of the training data by the long-term time series prediction model, calculate the loss score of the denormalized predicted data and the actual target time series data using a loss function, and train the long-term time series prediction model through backpropagation. Effects of the invention

[0024] The long-term time series prediction system according to the present invention, having the above-described configuration, can reduce the size of the model by using a long-term time series prediction model based on a neural network having a CNN-MLP hybrid structure, thereby replacing the Multi-Head Attention layer of the Transformer model with a single Linear layer.

[0025] In addition, the long-term time series forecasting system according to the present invention can enhance the processing capability for local information, where the Transformer model shows weaknesses, by using a convolutional layer. Brief explanation of the drawing

[0026] Figure 1 is a structural diagram illustrating an RNN neural network. Figure 2 is a structural diagram illustrating an LSTM neural network. Figure 3 is a structural diagram illustrating a TCN neural network. Figure 4 is a structural diagram illustrating a Transformer neural network. FIG. 5 is a configuration diagram illustrating a long-term time series prediction system according to a preferred embodiment of the present invention. FIG. 6 is a structural diagram illustrating a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the present invention. FIG. 7 is a structural diagram illustrating the data embedding section of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the present invention. FIG. 8 is a schematic diagram illustrating the time series patching process by the data embedding unit of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the present invention. FIG. 9 is a structural diagram showing the CNN-MLP hybrid mixer section of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the invention. FIG. 10 is a structural diagram illustrating the convolutional module of the CNN-MLP hybrid mixer section of the long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the invention. FIG. 11 is a structural diagram illustrating a convolutional mixing module of a CNN-MLP hybrid mixer section of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the invention. FIG. 12 is a structural diagram illustrating the Gated Attention module of the CNN-MLP hybrid mixer section of the long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the invention. Specific details for implementing the invention

[0027] The long-term time series prediction system according to the present invention is characterized by being composed of a long-term time series prediction model based on a neural network having a CNN-MLP hybrid structure. Hereinafter, a long-term time series prediction system and a method according to a preferred embodiment of the present invention will be described in detail with reference to the attached drawings.

[0028] FIG. 5 is a configuration diagram illustrating a long-term time series prediction system according to a preferred embodiment of the present invention. Referring to FIG. 5, the long-term time series prediction system (5) according to the present invention comprises a model learning device (50), a time series prediction device (60), and a database (70).

[0029] The above database (70) stores and manages time series data used as training data to train a long-term time series prediction model. The above model training device (50) trains and models the long-term time series prediction model described later using the training data of the above database. The above time series prediction device (60) is equipped with a time series data input unit (62), a long-term time series prediction model (64), and a prediction data output unit (66). The above time series prediction device performs time series data prediction using the long-term time series prediction model (64) trained by the above model training device, and outputs the predicted time series data.

[0030] FIG. 6 is a structural diagram illustrating a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the present invention. Referring to FIG. 6, the long-term time series prediction model (64) comprises a data embedding unit (82), a CNN-MLP hybrid mixer unit (84), and a head prediction unit (86) to predict and output time series data having a preset length for input time series data. The data embedding unit (82) patch embedding the time series data. The CNN-MLP hybrid mixer unit (84) divides the embedded data, performs linear transformation and convolution operations on each of the divided data, and combines and fuses the linearly transformed data and the convolutionally operated data. The head prediction unit (86) transforms the output of the CNN-MLP hybrid mixer unit into a one-dimensional form and then linearly transforms it to predict and output time series data having a target time series length (T). The structure and operation of each component of the long-term time series prediction model will be described in more detail below.

[0031] FIG. 7 is a structural diagram illustrating a data embedding unit of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the present invention. Referring to FIG. 7, the data embedding unit (82) comprises a normalization module (820), a patch module (822), and a first linear layer (824) to patch embedding the input data, which is the time series data, and output it.

[0032] The input data is univariate time series data of length L as, It is multivariate time series data having features. The normalization module (820) converts the features of the time series data into univariate time series data by sending them to the batch dimension, and performs normalization through RevIN. At this time, the shape of the time series data is at It is converted to. Here, None refers to the batch dimension, and in this specification It is written in a simplified form.

[0033] The patch module (822) divides the normalized time series data into patches of a preset length while allowing overlap. The length of the patch is P, and the stride is denoted as S. When applying padding to the time series data, the number of patches N is It becomes, and the shape is This becomes. FIG. 8 is a schematic diagram illustrating a time series patching process by a data embedding unit of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the present invention. Referring to FIG. 8, for time series data of length 10, the patch size is set to 4 and the stride is set to 2 to illustrate the patching process exemplarily.

[0034] The first linear layer (824) above uses the patch generated by the patch module as a first weight ( Using ) each patch pre-set latent dimension ( It is projected as ). The embedding data obtained by the first linear layer can be expressed as Equation 1.

[0035]

[0036] The above CNN-MLP hybrid mixer (84) divides the embedded data into first and second data, performs linear transformation and convolution operations on each of the divided data, and combines and fuses the linearly transformed data and the convolutionally operated data. In this way, the above CNN-MLP hybrid mixer (84) configures the linear layer and the convolutional layer in parallel, thereby reducing the model parameters compared to the Transformer's Self-Attention structure by using the linear layer, and supplementing the ability to process local information by using the convolutional layer.

[0037] FIG. 9 is a structural diagram illustrating the CNN-MLP hybrid mixer section of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the invention. Referring to FIG. 9, the CNN-MLP hybrid mixer section (84) comprises a data splitting module (840), a second linear layer (841), a convolutional module (842), a combination module (843), a convolutional mixing module (844), a Gated Attention module (845), and a Feed Forward network (846).

[0038] The above data partitioning module (840) is the patch embedded Input data in the form After Layer Normalization, from the latent dimension to the embedding dimension is the first dimension ( ) and the second dimension ( It is partitioned into ). The sum of the first and second dimensions is the embedding dimension. It must be either party. Input data is first-dimensional first data to be passed through a linear layer and second-dimensional second data to pass through the convolution module It is split. For example, input data with a batch size of 10, embedding dimension of 64, and number of patches of 48. In this case, the shape is (10, 64, 48). This is an embedding dimension of 16. Wow, the embedding dimension is 48 It will be divided into. At this time, and The shapes are (10,16,48) and (10,48,48), respectively.

[0039] The above second linear layer (841) is a second weight ( Using ), the first data is linearly transformed. Unlike the existing TSMixer-K model, the model according to the present invention has only a single linear layer, so computational complexity can be reduced. The value that has passed through the second linear layer ( ) can be expressed as in mathematical formula 2.

[0040]

[0041] FIG. 10 is a structural diagram illustrating a convolutional module of a CNN-MLP hybrid mixer section of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the invention. Referring to FIG. 10, the convolutional module comprises a third weight ( Using ) the above second data The convolution module (842) having the above configuration comprises a third linear layer that linearly transforms the second data, a depth-wise convolution layer that performs depth-wise convolution operations on the linearly transformed second data, and an activation layer that activates the second data convolutionally performed using an activation function. is the third weight ( It is configured to learn information between adjacent patches by passing through a third linear layer having ) and using depth-wise convolution, and then passing through an activation layer. The value that passed through the convolution module ( ) can be expressed as in mathematical formula 3.

[0042]

[0043] The second linear layer (841) and the convolutional module (842) are connected in parallel. Due to this parallel connection structure of the second linear layer (841) and the convolutional module (842), the first data among the input data that has undergone linear transformation learns the structure of the overall patches, and the second data among the input data that has undergone convolutional operation learns information between adjacent patches. In addition, by passing the second data that has undergone convolutional operation through a depth-wise convolution, it is possible to achieve an effect similar to using a general convolutional layer while simultaneously reducing the size of the model or the complexity of the calculation.

[0044] The above combining module (843) concatenates the linear data that has passed through the linear layer and the convolutional data that has passed through the convolutional module, It makes it into the form of.

[0045] FIG. 11 is a structural diagram illustrating a convolutional fusion module of the CNN-MLP hybrid mixer section of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the invention. Referring to FIG. 11, the convolutional fusion module (844; Conversion Fusion Module) combines information of data connected by the combination module (843) using depth-wise convolution. The convolutional fusion module (844) and It serves the role of combining the data divided into sections using depth-wise convolution. The value that has passed through the above convolution mixing module ( ) can be expressed as in mathematical formula 4.

[0046]

[0047] FIG. 12 is a structural diagram illustrating a Gated Attention module of a CNN-MLP hybrid mixer unit of a long-term time series prediction model in a long-term time series prediction system according to a preferred embodiment of the invention. Referring to FIG. 12, the Gated Attention module (845) [receives] data mixed by the convolutional mixing module as a fourth weight ( After linearly transforming through a fourth linear layer having ), output data is generated by applying a softmax function to the dimension axis between patches. Through this, the Attention structure is mimicked with a simple structure to help identify the correlation between each patch. The value that passed through the above Gated Attention module ( ) can be expressed as in mathematical formula 5.

[0048]

[0049] The above Feed Forward network (846) mixes information with the output data of the Gated Attention module along the latent dimension axis. The value that passes through the above Feed Forward network (846) ) can be expressed as in mathematical formula 6.

[0050]

[0051] The head prediction unit (86) above is the output of the CNN-MLP hybrid mixer unit. second After transforming into a 1-dimensional form, the 5th weight ( The time series data having a target time series length (T) is predicted and output by linearly transforming it through a fifth linear layer having ) ) can be expressed as in mathematical formula 7.

[0052]

[0053] Meanwhile, the above model learning device (50) uses time series data (in a CNN-MLP hybrid model) If ) is predicted, denormalization is performed on the predicted data using RevIN, and then the actual target time series ( After calculating the loss score using the ) and the MSE loss function, the model is trained through backpropagation.

[0054] The long-term time series forecasting model based on a neural network having a CNN-MLP hybrid structure according to the present invention, having the aforementioned configuration, has a parameter size of about 1 / 4 and a computational complexity of 1 / 6 compared to existing Transformer-based models. In addition, the model according to the present invention is a lightweight model that exhibits a smaller number of parameters and lower computational complexity compared to Transformer-based models, yet demonstrates superior MSE and MAE scores compared to Transformer-based models. Therefore, the model according to the present invention can be usefully employed in actual long-term time series forecasting fields, such as in environments where computing resources are limited or in situations where various time series forecasting models are executed simultaneously.

[0055] Although the present invention has been described above with reference to preferred embodiments, this is merely illustrative and does not limit the invention. Those skilled in the art will understand that various modifications and applications not exemplified above are possible within the scope of the essential characteristics of the invention. Furthermore, differences related to such modifications and applications should be interpreted as being included within the scope of the invention as defined in the appended claims. Explanation of the symbols

[0056] 5: Long-term Time Series Forecasting System 50: Model training device 60: Time series forecasting device 62: Long-term time series forecasting model 70 : Database 82 : Data Embedding Section 820 : Normalization module 822 : Patch Module 824: 1st Linear Layer 84 : CNN-MLP Hybrid Mixer Section 840: Data Partitioning Module 841: Second linear layer 842 : Convolution Module 843 : Combined Module 844 : Convolutional Mixing Module 845: Gated Attention Module 846 : Feed Forward Network 86 : Head prediction section

Claims

Claim 1 A long-term time series prediction device based on a CNN-MLP hybrid model, comprising: a data input unit for receiving time series data; a long-term time series prediction model modeled based on a neural network having a CNN-MLP hybrid structure; and a data output unit for outputting data predicted by the long-term time series prediction model; wherein the long-term time series prediction model comprises: a data embedding unit for patch embedding the input time series data; a CNN-MLP hybrid mixer unit for splitting the embedded data into first and second data, performing a linear transformation on the first data and a convolution operation on the second data, and combining and fusing the linearly transformed first data and the convolutionally performed second data; and a head prediction unit for transforming the output of the CNN-MLP hybrid mixer unit into one dimension, performing a linear transformation, and predicting and outputting time series data having a target time series length (T). Claim 2 A long-term time series forecasting device based on a CNN-MLP hybrid model according to claim 1, wherein the data embedding unit comprises: a normalization module that converts the time series data into univariate time series data and normalizes it; a patch module that divides the normalized time series data into patches having a preset length; and a first linear layer that projects each patch onto a preset latent dimension using a first weight, and outputs the input time series data by patch embedding. Claim 3 A long-term time series forecasting device based on a CNN-MLP hybrid model according to claim 1, wherein the CNN-MLP hybrid mixer comprises: a data splitting module that normalizes the patch-embedded data and then splits it into first data and second data in the latent dimension; a second linear layer that linearly transforms the first data using second weights; a convolution module that performs a convolution operation on the second data; a combination module that connects the first data linearly transformed by the second linear layer and the second data that has undergone a convolution operation by the convolution module; a convolution mixing module that combines information of the data connected by the combination module using depth-wise convolution; a Gated Attention module that linearly transforms the data mixed by the convolution mixing module and then applies a softmax function to generate output data; and a Feed Forward network that mixes information along the latent dimension axis for the output data of the Gated Attention module. Claim 4 A long-term time series forecasting device based on a CNN-MLP hybrid model, characterized in that, in claim 3, the convolution module comprises: a third linear layer that linearly transforms data using a third weight; a depth-wise convolution layer that performs depth-wise convolution operations on the data linearly transformed by the third linear layer; and an activation layer that activates the data convolutionally performed using an activation function, thereby being configured to learn information between adjacent patches. Claim 5 A long-term time series prediction device based on a CNN-MLP hybrid model according to claim 1, wherein the long-term time series prediction model is a model trained using training data including time series data and target time series data, wherein the data predicted from the time series data of the training data by the long-term time series prediction model is denormalized, the loss score of the denormalized predicted data and the actual target time series data is obtained using a loss function, and the long-term time series prediction model is trained through backpropagation. Claim 6 A long-term time series prediction method based on a neural network-based long-term time series prediction model, comprising: (a) a step of patch embedding input time series data; (b) a step of dividing the patch-embedding data into first and second data, then linearly transforming the first data and performing a convolution operation on the second data, and combining and fusing the linearly transformed data and the convolutionally performed data; and (c) a step of transforming the output of step (b) into one dimension and then linearly transforming it to predict and output time series data having a target time series length (T). Claim 7 A long-term time series forecasting method based on a CNN-MLP hybrid model, wherein, in claim 6, the step (a) comprises: (a1) converting the time series data into univariate time series data and normalizing it; (a2) dividing the normalized time series data into patches having a preset length; and (a3) ​​projecting each patch onto a preset latent dimension using a first weight, thereby embedding the input time series data into patches and outputting the result. Claim 8 In claim 6, the above step (b) comprises: (b1) a step of normalizing the patch-embedded data and then splitting it into first data and second data in the latent dimension; (b2) a step of linearly transforming the first data using second weights; (b3) a step of performing a convolution operation on the second data; (b4) a step of connecting the linearly transformed first data and the convolutionally generated second data; (b5) a step of combining the information of the data connected by step (b4) using depth-wise convolution; (b6) a step of linearly transforming the data mixed by step (b5) and then applying a softmax function to generate output data; and (b7) a step of mixing information along the latent dimension axis for the output data of step (b6) through a feed-forward network; characterized in that it is a long-term time series forecasting method based on a CNN-MLP hybrid model. Claim 9 A long-term time series forecasting method based on a CNN-MLP hybrid model, characterized in that, in claim 8, the step (b3) comprises: (b3-1) a step of linearly transforming data using a third weight; (b3-2) a step of performing a depth-wise convolution operation on the linearly transformed data; and (b3-3) a step of activating the convolution operation data using an activation function, thereby being configured to learn information between adjacent patches. Claim 10 In claim 6, the long-term time series prediction method further comprises: (d) a step of training a long-term time series prediction model based on a CNN-MLP hybrid model using training data including time series data and target time series data; wherein step (d) is characterized by denormalizing the data predicted from the time series data of the training data by the long-term time series prediction model, obtaining a loss score between the denormalized predicted data and the actual target time series data using a loss function, and training the long-term time series prediction model through backpropagation.

Images