Knowledge integration decomposition network for long-term time series prediction

Abstract

The invention discloses a knowledge integration decomposition network for long-term time sequence prediction, which comprises the following steps that: a multi-scale moving average module performs moving average on input time sequence data # imgabs0 # to obtain a trend sequence # imgabs1 # and a seasonal sequence # imgabs2 #, and connects # imgabs3 # and # imgabs4 # in parallel to obtain a multi-scale sequence # imgabs5 #; a season learning unit and a trend learning unit respectively learn a trend characteristic # imgabs7 # and a seasonal characteristic # imgabs8 # according to # imgabs6 #; a unit difference value # imgabs9 # is obtained according to the remainder items; the seasonal prediction unit is used for predicting a seasonal sequence # imgabs10 #; the trend prediction unit outputs a trend prediction sequence # imgabs11 #; the output unit is used for adding the # imgabs12 # and the # imgabs13 # to obtain a final prediction output # imgabs14 #; and the training module carries out joint training on the learning module and the prediction module.

Description

Technical Field

[0001] The present invention relates to the technical field of time series prediction, and in particular to a knowledge integration decomposition network for long-term time series prediction. Background Art

[0002] At present, many time series prediction models based on deep learning process historical data through sequence decomposition methods, decomposing complex time series into multiple components to extract temporal patterns and enhance interpretability. Among them, some models use moving average operations to extract trend features, but the moving average operation is less robust to noise and has limitations in capturing complex trends and seasonal patterns. In statistics, researchers have proposed many excellent sequence decomposition methods, such as STL, EMD, VMD, SSA, etc. Among them, the most widely used is STL decomposition, which can estimate nonlinear relationships, is highly robust to outliers and missing data, and can decompose time series of any seasonal frequency. However, directly applying these sequence decomposition methods based on statistical theory to deep learning models will cause the following problems: 1. Information leakage: In the past, models using STL decomposition performed an initial decomposition of the entire data set, and then divided all the resulting subsequences into three stages for training, validation, and testing. This approach may lead to a problem: during the testing phase, the model may "learn" the characteristics and patterns of future data, resulting in unrealistically high accuracy when predicting. This information leakage phenomenon seriously affects the generalization ability of the model and the authenticity of the prediction results.

[0003] 2. Poor decomposition effect of short sequences: To avoid information leakage, some models using STL and EMD decomposition adopt a step-by-step decomposition strategy. Although this step-by-step decomposition strategy reduces the risk of information leakage to a certain extent, it also has two problems. First, this method requires repeated decomposition operations on subsequences, which is not only time-consuming but also requires a lot of computing resources. Second, due to the short length of the subsequence, it cannot fully show the complete periodicity and trend characteristics, which makes it difficult to accurately capture key information during the decomposition process.

[0004] 3. Model consistency and optimization limitations: In existing studies, STL decomposition is often used as a preprocessing step for historical data. This separation of decomposition and prediction may lead to a lack of consistency in the model as a whole, affecting the overall optimization of the decomposition-prediction process. Especially when dealing with complex time series data, the disconnect between decomposition and prediction may affect the flexibility and accuracy of the model.

[0005] In order to overcome the above-mentioned problems, this paper proposes a knowledge-integrated decomposition network for long-term time series prediction, which integrates the sequence decomposition and prediction processes into a unified deep learning framework and uses a joint optimization strategy to simultaneously train the decomposition network and the prediction network. This method ensures that the two networks coordinate with each other during the training process and improves the accuracy of the prediction. Summary of the invention

[0006] In view of this, the present invention provides a knowledge integration decomposition network for long-term time series prediction, which is used to at least solve the problem in the prior art of how to integrate expert knowledge into deep neural networks to minimize the complexity of the model and improve performance.

[0007] In order to achieve the above object, the present invention adopts the following technical solution: A knowledge integration decomposition network for long-term time series prediction, comprising: a multi-scale moving average module, a learning module, a prediction module and a training module; The multi-scale moving average module is used to simultaneously select Convolution kernels of different sizes are used to compare time series Perform moving average to obtain Group trend series and seasonal sequence , and and Multi-scale sequence obtained in parallel ,in is the length of the time series, is the variable dimension; is the real number space; The learning module includes a season learning unit, a trend learning unit and a residual acquisition unit. The season learning unit and the trend learning unit are respectively used to obtain the multi-scale sequence Learning trend characteristics and seasonal characteristics ; The remaining item acquisition unit is used to obtain and and further calculate and sum and The difference between ; The prediction module includes a season prediction unit, a trend prediction unit and an output unit, wherein the season prediction unit is used to predict the season sequence The trend prediction unit is used to and Extracting local trend features and remainder features , and and After addition, the trend prediction sequence is output through linear transformation. The output unit is used to and Add together to get the final prediction output ; The training module is used to jointly train the learning module and the prediction module, and iteratively updates the model weights during the training process until the loss is minimized.

[0008] Preferably, the multi-scale moving average module converts the time series Decompose to get trend sequence and seasonal sequence The specific content is: ; ; in is to add zero elements to the sequence to maintain its length, is a moving average operation.

[0009] Preferably, the seasonal learning unit and the trend learning unit have the same structure, both including two two-dimensional convolutions, a fully connected layer and a linear layer; The two-dimensional convolution is used to fuse sequences from multiple scales. The information includes: ; The fully connected layer is used to obtain the fused features , used to capture time features, perform nonlinear mapping on time series, and change the length of time series from Map to ,in represents the mapping coefficient; The linear layer is used to convert the time series length after the full connection layer is mapped Linear mapping back : ; ; in, Indicates with The fully connected layer of the activation function, Represents a linear layer.

[0010] Preferably, the remainder acquisition unit is used to acquire the remainder , the specific contents are: Original sequence and and The difference between : ; in, and Respectively represent the decomposition module from the time series Learned trend, seasonality, and residual components.

[0011] Preferably, the season prediction unit includes a sequence embedding subunit, a detail feature extraction subunit, and a first feature fusion and prediction subunit; The sequence embedding subunit is used to fill it with zero values arrive and through Convolution transforms it from Dimensions are mapped to Dimensions, specifically: ; in Indicates the length is The zero matrix of ; The detail feature extraction subunit is used to use the convolution kernel size and step size of The one-dimensional convolution performs downsampling and obtains the sequence ;in Represents the extracted local features, different The values ​​correspond to multiple values ​​of different scales. ;right Apply one-dimensional convolution to learn local features. In order to maintain the original time granularity of the time series, Apply transposed convolution to restore it to the original sequence length, thereby obtaining the global features of the mapping ; The first feature fusion and prediction subunit is used to combine features from different branches Connect in parallel to obtain ,in is the number of branches; by applying Convolution, Aggregate the information in to obtain multi-scale seasonal features , using a linear layer from Dimensional mapping back Dimension, truncate the end of the sequence The subsequence of length is represented as the prediction result of the seasonal term .

[0012] Preferably, the trend prediction unit includes a trend feature extraction subunit, a residual feature extraction subunit, and a second feature fusion and prediction subunit; The trend feature extraction subunit includes two one-dimensional convolutional layers with different kernel sizes, specifically: ; ; in Represents the trend features learned by the convolution unit; The residual feature extraction subunit uses the highly nonlinear modeling capability of the fully connected layer to Nonlinear mapping to high-dimensional feature space and linear mapping back to the original sequence length, specifically: ; in is the residual feature, Indicates that Fully connected layer with activation function; The second feature fusion and prediction subunit is used to combine the trend feature and remainder feature Add and map through the linear layer to get the forecast sequence of trend and remainder .

[0013] Preferably, it also includes Decomposition module, the Decomposition module for time series conduct break down.

[0014] Preferably, the loss function in the training module is: ; ; ; in represents the decomposition loss, Include trending items , Seasonal items and remainder ; For the The decomposition module decomposes the obtained supervisory signal, including the corresponding trend component, seasonal component and residual component; is the prediction loss, which refers to the predicted value With the true value The difference between The training module uses The optimizer calculates and updates the gradient of the weight parameters to minimize the error. During the training process, the model weights are iteratively updated until the loss is minimized and the optimal prediction is achieved. The specific training algorithm is as follows: Input multivariate time series data , is the data length, is the number of variables; Set parameters, including: learning rate , batch size , input the history sequence length , predict sequence length And the number of training epochs; Cycle training network: (1) Initialize model parameters; (2) Input data into the model to obtain predicted values ; (3) Calculate the loss value loss through the loss function and update the model parameters with reverse gradient; (4) If the training number is reached, stop training; otherwise, repeat steps (2) to (4).

[0015] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a knowledge integration decomposition network for long-term time series prediction, which has the following beneficial effects: The network disclosed in the present invention mainly includes two parts: decomposition and prediction. The trend, season and residual components of the sequence are obtained through a multi-scale moving average module and a learning module. The residual represents the non-periodicity and short-term fluctuations of the time series. Therefore, decomposing and extracting features thereof helps the model to make more precise predictions. The three components output by the learning module represent the global time relationship of historical sequences of different scales, and feature extraction is performed on each component in the prediction module to generate a prediction. The final prediction result is the sum of the three items. It can be seen that the network combines decomposition with prediction with learnable parameters for collaborative training, thereby ensuring that the decomposition and prediction modules are coordinated with each other during the training process and improving the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A model structure diagram of a knowledge integration decomposition network for long-term time series prediction provided by the present invention; Figure 2A structural diagram of a multi-scale moving average module and a learning module in a knowledge integration decomposition network for long-term time series prediction provided by the present invention; Figure 3 A structural diagram of a prediction module in a knowledge integration decomposition network for long-term time series prediction provided by the present invention; Figure 4 A structural diagram of a seasonal prediction unit in a knowledge integration decomposition network for long-term time series prediction provided by the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The present invention provides a knowledge integration decomposition network for long-term time series prediction, such as Figure 1 As shown, it includes: a multi-scale moving average module, a learning module, a prediction module and a training module; Multi-scale moving average module, used to simultaneously select Convolution kernels of different sizes are used to compare time series Perform moving average to obtain Group trend series and seasonal sequence , and and Multi-scale sequence obtained in parallel ,in is the length of the time series, is the variable dimension; is the real number space; The learning module includes a seasonal learning unit, a trend learning unit and a residual acquisition unit. The seasonal learning unit and the trend learning unit are used to obtain the seasonality of the multi-scale sequence. Learning trend characteristics and seasonal characteristics ; The remainder acquisition unit is used to obtain the residual through the numerical estimation algorithm The generated values ​​are used as expert prior knowledge for supervised training to obtain and and further calculate and sum and The difference between ; The prediction module includes a season prediction unit, a trend prediction unit and an output unit, wherein the season prediction unit is used to predict the seasonal sequence. ; The trend prediction unit is used to and Extracting local trend features and remainder features , and and After addition, the trend prediction sequence is output through linear transformation. ; The output unit is used to and Add together to get the final prediction output ; The training module is used to jointly train the learning module and the prediction module, and iteratively update the model weights during the training process until the loss is minimized.

[0020] It should be noted that: Trend term and seasonal term can simply reflect the trend and seasonality of the sequence; complex time series contain a large amount of time dynamics. In order to make it easier for subsequent sequence learning units to learn the seasonal and trend characteristics of the time series, the present invention also selects The convolution kernels of different sizes are used for moving average.

[0021] In order to further implement the above technical solution, Figure 2 As shown, the multi-scale moving average module converts the time series Decompose to get trend sequence and seasonal sequence The specific content is: ; ; in , is to add zero elements to the sequence to maintain its length, is a moving average operation.

[0022] In order to further implement the above technical solution, Figure 2 As shown in Figure 2, the seasonal learning unit and the trend learning unit have the same structure, both of which include two 2D convolutions, a fully connected layer, and a linear layer. 2D convolution is used to fuse sequences from multiple scales The information includes: ; The fully connected layer is used to obtain the fused features , used to capture time features, perform nonlinear mapping on time series, and change the length of time series from Map to ,in represents the mapping coefficient; The linear layer is used to map the time series length after the fully connected layer Linear mapping back : ; ; in, Indicates with The fully connected layer of the activation function, Represents a linear layer.

[0023] It should be noted that: Multi-scale sequence are input into two learning blocks with the same structure to learn trend and seasonal features independently; to ensure that the two learning units can learn trend features separately and seasonal characteristics , the present invention uses the classic numerical estimation algorithm The generated values ​​are used as expert prior knowledge for supervised training.

[0024] In order to further implement the above technical solution, the remainder acquisition unit is used to acquire the remainder , the specific contents are: Original sequence and and The difference between : ; in, and Respectively represent the decomposition module from the time series Learned trend, seasonality, and residual components.

[0025] In order to further implement the above technical solution, Figure 3-4 As shown, the season prediction unit includes a sequence embedding subunit, a detail feature extraction subunit, and a first feature fusion and prediction subunit; Sequence embedding subunit for padding with zero values arrive and through Convolution transforms it from Dimensions are mapped to Dimensions, specifically: ; in Indicates the length is The zero matrix of ; The detail feature extraction subunit is used to use the convolution kernel size and step size The one-dimensional convolution performs downsampling and obtains the sequence ;in Represents the extracted local features, different The values ​​correspond to multiple values ​​of different scales. ;right Apply one-dimensional convolution to learn local features. In order to maintain the original time granularity of the time series, Apply transposed convolution to restore it to the original sequence length, thereby obtaining the global features of the mapping ; The first feature fusion and prediction subunit is used to combine features from different branches Connect in parallel to obtain ,in is the number of branches; by applying 2D convolution, Aggregate the information in to obtain multi-scale seasonal features , using a linear layer from Dimensional mapping back Dimension, truncate the end of the sequence The subsequence of length is represented as the prediction result of the seasonal term .

[0026] It should be noted that: The seasonal prediction unit uses the detail feature extraction unit to extract seasonal features of different scales for the seasonal component, and then aggregates these multi-scale seasonal features through 2D convolution. In the sequence embedding subunit, in order to adapt to the prediction length , first fill the seasonal series with zero values arrive , and map it from D dimensions to N dimensions through 1D convolution.

[0027] In order to further implement the above technical solution, the trend prediction unit includes a trend feature extraction subunit, a residual feature extraction subunit and a second feature fusion and prediction subunit; The trend feature extraction subunit includes two one-dimensional convolutional layers with different kernel sizes, specifically: ; ; in Represents the trend features learned by the convolution unit; The residual feature extraction subunit uses the highly nonlinear modeling capability of the fully connected layer to Nonlinear mapping to high-dimensional feature space and linear mapping back to the original sequence length, specifically: ; in is the residual feature, Indicates that Fully connected layer with activation function; The second feature fusion and prediction subunit is used to transform the trend features and remainder feature Add and map through the linear layer to get the forecast sequence of trend and remainder .

[0028] It should be noted that: The trend feature extraction subunit extracts the detailed trend features of the trend component through the convolution unit. The residual feature extraction subunit, the residual component obtained from the multi-scale moving average module shows irregular fluctuations. In order to learn the residual features, the present invention uses the highly nonlinear modeling ability of the fully connected layer to Nonlinear mapping to high-dimensional feature space, and then linear mapping it back to the original sequence length. In order to improve the accuracy of trend sequence prediction and suppress the accumulation of prediction errors, this method will learn the trend features and remainder feature Then, the trend and remainder forecast sequence is obtained through linear layer mapping .

[0029] In order to further implement the above technical solution, it also includes Decomposition modules, Decomposition module for time series conduct break down.

[0030] In order to further implement the above technical solution, the loss function in the training module is: ; ; ; in represents the decomposition loss, Derived from the multi-scale moving average module, including trend term , Seasonal items and remainder ; For the The decomposition module decomposes the obtained supervisory signal, including the corresponding trend component, seasonal component and residual component; is the prediction loss, which refers to the predicted value With the true value The difference between During the neural network training process, the training module uses the Adam optimizer to calculate and update the gradient of the weight parameters to minimize the error. During the training process, the model weights are iteratively updated until the loss is minimized and the optimal prediction is achieved. The specific training algorithm is as follows: Input multivariate time series data , is the data length, is the number of variables; Set parameters, including: learning rate , batch size , input the history sequence length , predict sequence length And the number of training epochs.

[0031] Cycle training network: (1) Initialize model parameters; (2) Input data into the model to obtain predicted values ; (3) Calculate the loss value loss through the loss function and update the model parameters with reverse gradient; (4) If the training number is reached, stop training; otherwise, repeat steps (2) to (4).

[0032] It should be noted that: During the neural network training process, the Adam optimizer is used to calculate and update the gradient of the weight parameters to minimize the error. The advantage of Adam is that it can provide an independent adaptive learning rate for each parameter, thereby improving computational efficiency. During the training process, the model weights are iteratively updated until the loss is minimized and the optimal prediction is achieved.

[0033] The present invention will be further described below through experiments: The network model disclosed in the present invention is experimented on 8 mainstream time series prediction datasets, which cover multiple fields such as energy, transportation and weather, as shown in Table 1; the following is a brief description of these experimental datasets: (1) ETT contains two hourly datasets (ETTh) and two 15-minute datasets (ETTm); each dataset records seven oil load characteristics of power transformers from July 2016 to July 2018; (2) Electricity records hourly electricity consumption data of 321 users from 2012 to 2014; (3) Weather collects 21 10-minute meteorological indicators such as humidity and temperature in a certain country in 2020; (4) PEMS08 collects traffic flow data on a certain highway from July 1, 2016 to August 31, 2016; (5) Solar records the solar power generation of 137 photovoltaic power stations in a certain place in 2016. All datasets are divided into training set, validation set and test set in chronological order. Among them, the division ratio of ETT dataset is 6:2:2, and that of other datasets is 7:1:2.

[0034] ; Baseline: We selected 5 models with the highest accuracy from the existing technologies, 2 non-Transformer based models including MICN and DLinear, and 3 Transformer based models including PatchTST, FEDformer, and Autoformer as baselines.

[0035] Experimental setup: Set the prediction length for all datasets The look-back window L of FEDformer, Autoformer, and Informer is set to 96, and the look-back window L of other models is set to 336. In the univariate prediction, the OT variable in the data set is selected as the predictor variable.

[0036] Multivariate Experiment Results The multivariate prediction results are shown in Table 2, in which the best results are highlighted in bold and the second best results are underlined. In the table, KIDNet represents the network model proposed by the present invention. In multivariate long-term prediction, the model of the present invention outperforms all baseline methods. In 32 mean absolute error comparisons on 8 benchmark data sets, the model of the present invention achieved the best 25 times. And compared with the currently most accurate PatchTST model, the mean absolute error of the model of the present invention on the ETTm1 data set is reduced by 2.8%, the mean absolute error on the ETTm2 data set is reduced by 4.1%, the mean absolute error on the PEMS08 data set is reduced by 6.5%, and the mean absolute error on the Solar data set is reduced by 19.1%.

[0037] ; Univariate Experiment Results The univariate prediction results are shown in Table 3. In the 24 mean absolute error comparisons with all baseline methods, the proposed model achieved the best results 16 times. Compared with the PatchTST model, the proposed model reduced the mean absolute error by 7.7% on the ETTh2 dataset, 4.3% on the ETTm2 dataset, and 5.6% on the Electricity dataset.

[0038] ; The present invention proposes a knowledge integration decomposition network for long-term time series prediction, constructs an end-to-end sequence decomposition and prediction framework, and improves the prediction performance for long-term time series.

[0039] By adopting a joint optimization strategy, the decomposition module is embedded in the deep learning model, so that the decomposition and prediction modules can be trained together and share information. This training method enhances the model's ability to understand the characteristics and rules of time series data, and improves the model's training effect and prediction performance.

[0040] By using expert prior knowledge to assist in training the decomposition module, this method can effectively guide the network to learn the implicit patterns and features of time series, improve the model's fitting ability and enhance its interpretability.

[0041] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A knowledge integration decomposition network for long-term time series prediction, characterized in that: include: Multi-scale moving average module, learning module, prediction module and training module; The multi-scale moving average module is used to simultaneously select Convolution kernels of different sizes are used to compare the input time series Perform moving average to obtain Group trend series and seasonal sequence , and and Multi-scale sequence obtained in parallel ,in is the length of the time series, is the variable dimension; is the real number space; The learning module includes a season learning unit, a trend learning unit and a residual acquisition unit. The season learning unit and the trend learning unit are respectively used to obtain the multi-scale sequence Learning trend characteristics and seasonal characteristics ; The remaining item acquisition unit is used to obtain and and further calculate and sum and The difference between ; The prediction module includes a season prediction unit, a trend prediction unit and an output unit, wherein the season prediction unit is used to predict the season sequence The trend prediction unit is used to and Extracting local trend features and remainder features , and and After addition, the trend prediction sequence is output through linear transformation. The output unit is used to and Add together to get the final prediction output ; The training module is used to jointly train the learning module and the prediction module, and iteratively updates the model weights during the training process until the loss is minimized.

2. The knowledge integration decomposition network for long-term time series prediction according to claim 1, characterized in that: The multi-scale moving average module converts the time series Decompose to get trend sequence and seasonal sequence The specific content is: ； ； in is to add zero elements to the sequence to maintain its length, is a moving average operation.

3. The knowledge integration decomposition network for long-term time series prediction according to claim 1, characterized in that: The seasonal learning unit and the trend learning unit have the same structure, both including two two-dimensional convolutions, a fully connected layer and a linear layer; The two-dimensional convolution is used to fuse sequences from multiple scales. The information includes: ； The fully connected layer is used to obtain the fused features , used to capture time features, perform nonlinear mapping on time series, and change the length of time series from Map to ,in represents the mapping coefficient; The linear layer is used to convert the time series length after the full connection layer is mapped Linear mapping back : ； ； in, Indicates with The fully connected layer of the activation function, Represents a linear layer.

4. The knowledge integration decomposition network for long-term time series prediction according to claim 1, characterized in that: The remainder acquisition unit is used to obtain the remainder , the specific contents are: Original sequence and and The difference between : ； in, and .

5. The knowledge integration decomposition network for long-term time series prediction according to claim 1, characterized in that: The seasonal prediction unit includes a sequence embedding subunit, a detail feature extraction subunit, and a first feature fusion and prediction subunit; The sequence embedding subunit is used to fill it with zero values arrive and through Convolution transforms it from Dimensions are mapped to Dimensions, specifically: ； in Indicates the length The zero matrix of ; The detail feature extraction subunit is used to extract Use convolution kernel size and stride as The one-dimensional convolution performs downsampling and obtains the sequence ;in Represents the extracted local features, different The values ​​correspond to multiple values ​​of different scales. ;right Apply one-dimensional convolution to learn local features. In order to maintain the original time granularity of the time series, Apply transposed convolution to restore it to the original sequence length, thereby obtaining the global features of the mapping ; The first feature fusion and prediction subunit is used to combine features from different branches Connect in parallel to obtain ,in is the number of branches; by applying Convolution, Aggregate the information in to obtain multi-scale seasonal features , using a linear layer from Dimensional mapping back Dimension, truncate the end of the sequence The subsequence of length is represented as the prediction result of the seasonal term .

6. The knowledge integration decomposition network for long-term time series prediction according to claim 1, characterized in that: The trend prediction unit includes a trend feature extraction subunit, a residual feature extraction subunit and a second feature fusion and prediction subunit; The trend feature extraction subunit includes two one-dimensional convolutional layers with different kernel sizes, specifically: ； ； in Represents the trend features learned by the convolution unit; The residual feature extraction subunit uses the highly nonlinear modeling capability of the fully connected layer to Nonlinear mapping to high-dimensional feature space and linear mapping back to the original sequence length, specifically: ； in is the residual feature, FC represents the fully connected layer with ReLU activation function; The second feature fusion and prediction subunit is used to combine the trend feature and remainder feature Add and map through the linear layer to get the forecast sequence of trend and remainder .

7. The knowledge integration decomposition network for long-term time series prediction according to claim 1, characterized in that: Also included is an STL decomposition module, which is used for time series Perform STL decomposition.

8. The knowledge integration decomposition network for long-term time series prediction according to claim 7, characterized in that: The loss function in the training module is: ； ； ； in represents the decomposition loss, Include trending items , Seasonal items and remainder ; The supervisory signal decomposed by the STL decomposition module includes a corresponding trend component, a seasonal component and a residual component; is the prediction loss, which refers to the predicted value With the true value The difference between During the neural network training process, the training module uses the Adam optimizer to calculate and update the gradient of the weight parameters to minimize the error. During the training process, the model weights are iteratively updated until the loss is minimized and the optimal prediction is achieved. The specific training algorithm is as follows: Input time series data , is the data length, is the number of variables; Set parameters, including: learning rate , batch size , input the history sequence length , predict sequence length And the number of training epochs; Cycle training network: (1) Initialize model parameters; (2) Input data into the model to obtain predicted values ; (3) Calculate the loss value loss through the loss function and update the model parameters with reverse gradient; (4) If the training number is reached, stop training; otherwise, repeat steps (2) to (4).

Images