Long-term series prediction methods based on channel independence and bidirectional normalization

Abstract

This invention discloses a long-term time series prediction method based on channel independence and bidirectional normalization. The steps are as follows: constructing a bidirectional normalized learnable subnetwork; constructing a channel-independent dual-path prediction subnetwork; constructing a complete long-term time series prediction network; constructing a multivariate time series prediction dataset; setting up a training environment and training the long-term time series prediction network; and predicting the long-term time series. The bidirectional normalized learnable subnetwork of this invention can effectively handle the distribution displacement problem between input sequences and between input and output. Through the channel-independent dual-path prediction subnetwork, it solves the problem of inter-channel distribution displacement in multivariate time series, improving the prediction accuracy and stability of the model. This invention is applicable to long-term time series prediction in multiple fields such as power, weather, and transportation.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and more specifically relates to a long-term series forecasting method based on channel independence and bidirectional normalization within the field of electronic data processing technology. This invention can be widely applied to long-term series forecasting scenarios such as power load forecasting, traffic flow forecasting, weather forecasting, and economic forecasting. Background Technology

[0002] Deep neural networks play a crucial role in various applications of long-term time series forecasting (LTSF), such as power generation forecasting, traffic flow forecasting, weather forecasting, and economic forecasting. However, time series data often exhibits changes in statistical distribution over time, a phenomenon known as "distribution shift." This problem causes inconsistencies in data distribution between the training and testing phases of deep learning models, affecting the accuracy and stability of predictions, thus posing a significant challenge to deep learning-based long-term forecasting. In long-term time series forecasting tasks, distribution shift manifests in three main aspects: First, distribution shift between input sequences: the statistical distribution of input data may change significantly at different points in the same time series. For example, in power load forecasting, winter and summer power demand patterns are typically different, leading to inconsistencies between the distribution of training and forecast data. Second, distribution shift between input and output: the relationship between the input sequence and the forecast target can also shift due to changes in the external environment. For instance, in weather forecasting, the statistical relationship between historical weather data and predicted future weather conditions may change due to climate change or unforeseen events. Thirdly, there is the distribution shift between channels: In multivariate time series, the statistical relationships between different variables may change over time. This covariate shift is particularly common in multi-channel time series forecasting, and traditional mixed-channel processing methods often struggle to cope with such complex distribution changes.

[0003] Traditional time series forecasting methods, such as the Autoregressive Integrated Moving Average (ARIMA) model, primarily rely on the statistical characteristics of historical data and assume that the data distribution remains relatively stable within the forecast period. However, real-world time series data often exhibit dynamic changes in distribution due to external environmental variations, leading to poor performance of traditional methods when dealing with distribution shifts. With the development of deep learning technologies, models such as Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and Transformer-based models have been increasingly applied to time series forecasting. However, these methods, when performing long-term series forecasting, do not address the distribution shift problem at the model level. Therefore, they still face issues such as decreased prediction accuracy and model instability caused by distribution shifts.

[0004] Kim et al., in their paper "Reversible instance normalization for accurate time-series forecasting against distribution shift" (International Conference on Learning Representations (ICLR), 2021), proposed a learnable normalization method for time series forecasting based on a simple linear network. This method employs a two-stage approach. First, the input sequence is processed by a linear neural network to obtain fluctuation and scaling terms, which are then normalized to transform the input sequence into a uniform distribution space. Second, after prediction by the time series forecasting network, the model output is inversely normalized using the previously obtained fluctuation and scaling terms, transforming it back to the original space. This method allows different time series to complete model computation within the same distribution space through nonlinear transformation, alleviating the problem of uneven distribution between different input sequences. However, this method still has limitations. The simple linear network used has limited performance in handling complex nonlinear data, and it only adaptively scales the input sequence, thus failing to effectively address the distribution shift problem between input and output, as well as between channels.

[0005] Hangzhou Dianzi University disclosed a multi-step time series prediction method in its patent application, "A Time Series Prediction Method Based on Multivariate Long-Term Time Series Analysis" (Application No.: CN 202410603210.3, Publication No.: CN 118349806 A). The method involves several steps: First, data preprocessing and segmentation are performed, setting the lengths of historical and predicted sequences according to different application scenarios, and dividing the data into training and testing sets. Second, a statistical linear fitting network is pre-trained on the training data, and then a biased generative network, a multi-dimensional graph learning network, and a dual-dataflow learning network are combined for feature learning and prediction, ultimately generating the predicted sequence. This method, through the multi-dimensional graph learning network and the dual-dataflow learning network, can better capture the complex relationships in time series and uses the biased generative network to fit statistical values, addressing the distribution shift problem based on RevIN. However, a remaining drawback is the need for pre-training the linear fitting network, making the process relatively complex. Meanwhile, this method only addresses the scaling of distribution displacement between input sequences, and still fails to effectively solve the problem of distribution displacement between input and output, as well as between channels, in multivariate time series. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art by proposing a long-term time series prediction method based on channel independence and bidirectional normalization, which aims to solve the problems of decreased prediction accuracy and training instability caused by three types of distribution shifts in long-term time series prediction.

[0007] To achieve the above objectives, this invention introduces a novel model architecture combining a bidirectionally normalized learnable subnetwork and a channel-independent dual-path prediction subnetwork. The bidirectionally normalized learnable subnetwork employs two structurally identical multilayer perceptron modules to predict the fluctuation and scale terms of the input and output sequences, respectively, and performs normalization and inverse normalization on the input and output sequences, effectively addressing the distribution shift problem between input sequences and between input and output sequences. This invention introduces a channel-independent dual-path prediction subnetwork, employing a dual-path approach: one path independently predicts the patterns of each channel, while the other uses a sparse attention mechanism to fuse information between channels. This adaptive fusion addresses the distribution shift problem between channels and improves feature capture of relevant channels. During training, this invention utilizes an adaptive adjustment strategy for model parameters to effectively address the distribution shift problem in long-term time series prediction, thereby improving the accuracy and stability of long-term time series prediction.

[0008] This invention includes the following steps:

[0009] Step 1: Build two multilayer perceptron modules with identical structures. Set one multilayer perceptron module as a forward normalization learning module and the other multilayer perceptron module as a reverse normalization learning module to obtain a bidirectional normalized learnable subnetwork.

[0010] Step 2: After building the multilayer perceptron prediction module and the sparse attention prediction module in parallel, they are then connected in series with the dual-output fusion module to form a dual-path prediction sub-network based on channel independence.

[0011] Step 3: Connect the forward normalization learning module, the channel-independent dual-path prediction sub-network, and the inverse normalization learning module in series to form a long-term series prediction network.

[0012] Step 4: Use the sliding window segmentation method to segment the historical data S = {s} for predicting future time series. i |i∈[1,T],s i ∈R 1×m}, thus obtaining a multivariate time series dataset X = {x1, x2, ..., x} n The dataset is divided into training, validation, and test sets in chronological order, where m is the dimension of the historical data, and s... i Let T be the multidimensional historical data at time i, where T is the length of the historical data sequence and n is the total number of samples in the multivariate time series dataset.

[0013] Step 5: Adaptively adjust network parameters:

[0014] Set up the training environment for the long-term series prediction network, input the training set into the long-term series prediction network for computation, use the gradient function to iteratively update the network parameters, and fine-tune the model using the validation set until the network loss function converges. Evaluate the model's performance metrics using the test set to obtain the trained time series prediction network.

[0015] Step 6: Using the same method as in Step 4, the data of the long-term series to be predicted is segmented by sliding window and then input into the trained time series prediction network to output the prediction result.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] First, this invention effectively solves the problem of distribution shift between input sequences and between input and output in long-term time series prediction by using a bidirectionally normalized learnable subnetwork and adaptively adjusting the parameters of normalization and inverse normalization. This allows the invention to significantly improve the prediction accuracy and stability of the model while ensuring data scale consistency, reducing reliance on manually set parameters and lowering the model's sensitivity to changes in data distribution.

[0018] Secondly, this invention successfully addresses the distribution shift problem between channels in multivariate time series prediction by employing a channel-independent dual-path prediction subnetwork. One path, a channel-independent multilayer perceptron prediction module, enables more accurate feature extraction for each channel; the other path, a sparse attention prediction module, effectively integrates information from different channels by focusing on highly correlated channel features. This improves the overall accuracy of the prediction results while reducing computational complexity, thus adapting to the processing needs of high-dimensional data.

[0019] Third, the present invention can effectively address the distribution shift problem in long-term time series prediction through an adaptive adjustment strategy for model parameters, thereby significantly improving the accuracy and robustness of the prediction model. It has significant application advantages, especially in complex application scenarios involving multi-channel time series. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention;

[0021] Figure 2 This is a framework diagram of the dual-path prediction subnetwork of the present invention;

[0022] Figure 3 This is a framework diagram of the long-term series prediction model of the present invention;

[0023] Figure 4 , Figure 5 , Figure 6 The figures show the results of three simulation experiments of this invention. Detailed Implementation

[0024] The specific embodiments and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0025] Reference Figure 1 The implementation steps of the embodiments of the present invention will be further described in detail below.

[0026] Step 1: Build two multilayer perceptron modules with identical structures. Set one multilayer perceptron module as the forward normalization learning module and the other multilayer perceptron module as the inverse normalization learning module to obtain a bidirectional normalized learnable subnetwork.

[0027] Both the forward normalization learning module and the inverse normalization learning module are composed of a first linear layer, an activation layer, and a second linear layer connected in series. The input dimensions of the first linear layer are L and H, respectively, and the output is set to 64. The activation layer is implemented using the ReLU function. The input dimension of the second linear layer is set to 64, and the output size is 2. The parameters and bias of the second linear layer are initialized to 0, where the value of L is equal to the length of the historical time series of a sample, and the value of H is equal to the length of the future time series to be predicted.

[0028] Since time series data may exhibit significant distributional differences between different time periods, this distributional shift can affect the model's learning and predictive capabilities. Introducing learnable normalization parameters allows the model to automatically adjust the normalization process based on data characteristics, thereby better adapting to different input data. Furthermore, bidirectional normalization methods can not only alleviate the distributional shift problem between different inputs but also solve the distributional shift problem between input and output.

[0029] Step 2: After building the multilayer perceptron prediction module and the sparse attention prediction module in parallel, they are then connected in series with the dual-output fusion module to form a channel-independent dual-path prediction subnetwork.

[0030] according to Figure 2 The diagram shown illustrates the framework of the dual-path prediction subnetwork, and provides a more detailed description of how the various modules implemented by the dual-path prediction subnetwork are combined to form a complete long-term series prediction model.

[0031] The aforementioned channel-independent multilayer perceptron prediction module consists of parallel linear layers with the same number of layers as the historical data dimension. The input dimension of each linear layer is equal to the length L of the historical time series of the sample, and the output dimension is equal to the length H of the predicted future time series. This module processes each channel independently to fully capture the feature information of each channel, avoid interference, and make feature extraction for each channel more accurate.

[0032] The specific implementation of the channel independence processing step is as follows: model each channel separately, assuming the input channels of the multivariate time series are... For each channel Independently construct model f i Each feature is represented as follows:

[0033]

[0034] in, Indicates channel Its characteristics.

[0035] The sparse attention prediction module is structured as follows: a linear transformation layer, an encoder attention layer, and a linear projection layer. The input dimension of the linear transformation layer is set to be equal to the length L of the historical time series, and the output dimension is set to 512. The attention layer scaling factor (dropout) is set to 0.3, the hidden layer dimension is set to 2048, and the activation function is implemented using the GELU function. The input dimension of the linear projection layer is set to 512, and the output dimension is equal to the length H of the predicted future time series. The sparse attention prediction module compensates for the shortcomings of channel-independent multilayer perceptron prediction modules in fully utilizing information from other channels by focusing only on the most relevant inter-channel information, while maintaining the ability to capture important information.

[0036] The dual-output fusion module has a structure of [m, 2] learnable parameters, where m is the dimension of the historical data. The parameters of the learnable parameter layer of the dual-output fusion module are all set to 0.5. In high-dimensional multi-channel data, the number of channels is large and their interrelationships are complex. Simple independent channel subnetworks may ignore the collaborative information between channels, while simple fusion processing may lead to excessive computation or over-fusion of information. Therefore, the combination of independent channel processing and sparse attention mechanisms can selectively fuse the correlations between key channels while capturing independent channel information, thus addressing the complexity of the data.

[0037] Step 3, according to Figure 3 The model structure shown below combines the various modules implemented in the preceding process into a complete long-term series prediction model, which will be described in further detail. Specifically, the forward normalization learning module, the channel-independent dual-path prediction sub-network, and the inverse normalization learning module are sequentially connected in series to form a long-term series prediction network.

[0038] Step 4, use the historical data S = {s} to predict future time series data. i |i∈[1,T],s i ∈R 1×m Using sliding window segmentation, a multivariate time series dataset X = {x1, x2, ..., x...} is constructed. n The dataset is divided into training, validation, and test sets in chronological order, where m is the dimension of the historical data, and s is the dimension of the test set. i Let T be the multidimensional historical data at time i, where T is the length of the historical data sequence and n is the total number of samples in the multivariate time series dataset.

[0039] Dividing time series data into smaller time windows allows the model to focus on capturing local time series patterns and trends, rather than trying to capture all information across the entire series. This helps to assume that the data is stationary within each time window, making the modeling process simpler and more effective. Therefore, in time series analysis and forecasting tasks, continuous time series data are often divided into multiple time windows for processing. The specific steps of data processing are: given a time series S = {s...} i |i∈[1,T],s i ∈R 1×m}, where T is the total length of the sequence and m is the dimension of the variable. Data segmentation is performed using a fixed scrolling window of length L+H to obtain... Where n represents the number of samples after segmentation, the data in a window sequence is divided into two parts: the input sequence, i.e., the historical observations X. t ={s t-L+1 ,...,s t The output sequence is the predicted value Y. t ={s t+1 ,...,s t+H} where L is the length of the historical data in a sample, i.e., the input length, and H is the length of the predicted future sequence. The dataset is divided into training, validation, and test sets in a chronological order of 8:1:1.

[0040] Step 5, adaptively adjust network parameters: Set up the training environment for the long-term series prediction network, input the training set into the long-term series prediction network for computation, use the gradient function to iteratively update the network parameters, and fine-tune the model using the validation set until the network loss function converges. Evaluate the model's performance metrics using the test set to obtain the trained time series prediction network.

[0041] The training environment for the long-term series prediction network is as follows: the network loss function is the mean squared error (MSE), the optimizer is Adam, the initial learning rate is set to 0.001, the bath size is set to 32, and the early stop during training is set to 10 epochs.

[0042] The computation steps of the long-term series prediction network are as follows:

[0043] The first step involves inputting the original multidimensional time series samples X from the prediction task into the long-term series prediction network. Through a forward normalization learning module, the input original multidimensional time series... Normalization is performed to eliminate the distribution shift between the original input sequences, resulting in a multidimensional time series with input offset repair, as shown in the following expression:

[0044]

[0045] Where X represents the input time series data, μ and σ are the mean and standard deviation, respectively, γ and β are learnable scaling and offset parameters, and ⊙ represents element-wise multiplication.

[0046] The second step involves passing the offset-corrected multidimensional time series data through a channel-independent dual-path prediction subnetwork. One path is input into a channel-independent multilayer perceptron prediction module, and each channel is input into its corresponding multilayer perceptron to extract the independent feature representation of each channel, i.e., the channel-independent result matrix, as shown in the following expression:

[0047]

[0048]

[0049] Where i represents the i-th dimension, f i This represents a multilayer perceptron with the i-th dimension. Let i be the feature vector of the i-th channel. This is the result matrix for channel-independent operation.

[0050] Another approach involves using a sparse attention network prediction module to capture the correlation and covariate information between channels in each channel sequence, thus obtaining the results of the sparse attention mechanism. The result matrix of multi-channel attention prediction is expressed as follows:

[0051]

[0052] Where Embedding(·) represents the computation of the linear transformation layer, attention(·) represents the computation of the sparse attention layer, and Projection(·) represents the computation of the mapping layer. This is the result matrix for multi-channel attention prediction.

[0053] Third, the channel-independent result matrix and the multi-channel attention prediction result matrix are input together into the dual-output fusion module to obtain the dual-fusion result matrix, as shown in the following expression:

[0054]

[0055] Where merge(·) represents the calculation of dual-path fusion. This represents the result matrix of the dual-path fusion;

[0056] The fourth step involves inverse normalization of the dual-path fusion result matrix via an inverse normalization learning module to restore the original scale of the prediction results and eliminate the distribution shift between the input and output. The expression is as follows:

[0057]

[0058] Where μ' and σ' are the learnable mean and standard deviation, respectively, and γ' and β' are the learnable scaling and offset parameters, resulting in the final multidimensional time series prediction for the future.

[0059] The network parameters are iteratively updated using a gradient function, with the model being fine-tuned on a validation set until the network loss function converges. The model's performance metrics are then evaluated using a test set, resulting in a well-trained time series prediction network. The network loss function is as follows:

[0060]

[0061] Where Y represents the true future sequence of the training set samples. Let H represent the future sequence predicted by the long-term time series prediction model, m represent the total number of channels, H represent the length of the predicted future sequence, i represent the i-th channel, and j represent the j-th time point of the future sequence.

[0062] Step 6: Using the same method as in Step 4, the data of the long-term series to be predicted is segmented by sliding window and then input into the trained time series prediction network to output the prediction result.

[0063] The effectiveness of this invention can be further demonstrated through the following simulation.

[0064] 1. Simulation experimental conditions.

[0065] The simulation platform for this invention is: an Intel(R) Core(TM) i7-10700K CPU with a main frequency of 3.80GHz, 64.0GB of memory, a GeForce RTX 3090*1 graphics card, an Ubuntu 22.04 operating system, and a PyCharm2024.2 development platform.

[0066] The simulation experiments of this invention were conducted on the publicly available time-series forecasting datasets Electricity, Weather, Traffic, and ETT. The Electricity dataset, published by Portuguese researcher Artur Trindade, contains hourly electricity consumption data from 321 customers collected from 2012 to 2014. The Weather dataset collects data on 21 weather indicators, such as air temperature and humidity, recorded every 10 minutes in a region of Germany in 2020. The Traffic dataset collects hourly data from the California Department of Transportation for 48 months (2015-2016), describing road occupancy measured by different sensors on highways in the San Francisco Bay Area. These datasets are all publicly available, multi-dimensional time-series datasets from their respective fields and are widely used in time-series forecasting research.

[0067] 2. Simulation content and result analysis.

[0068] The simulation experiments of this invention include three simulations respectively based on the Traffic dataset, the Weather dataset, and the Electricity dataset.

[0069] 2.1 Simulation Experiment 1 is a simulation of the Traffic dataset.

[0070] Simulation Experiment 1 of this invention uses the method of this invention and two existing technologies: the Informer algorithm and the DLinear algorithm, to simulate the Traffic dataset. It obtains time series prediction results for historical input sequence lengths of 336 and predicted future sequence lengths of 96, 192, 336, and 720, respectively. These results are compared with real data to obtain the MSE (Mean Sequence Equation). The obtained scores are then plotted as shown in the figure. Figure 4 The bar chart shown.

[0071] In simulation experiment 1, the two existing technologies used specifically refer to:

[0072] The existing technology, the Informer algorithm, is a long-sequence forecasting method proposed by Zhou H et al. in their paper "Informer: Beyond efficient transformer for long sequence time-series forecasting" (The Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI) , pp 11106-11115, 2021), abbreviated as Informer algorithm.

[0073] The existing technology, DLinear algorithm, is a long-term time series forecasting method proposed by Zeng A et al. in their paper "Are transformers effective for time series forecasting?" (The Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI) , 2023), abbreviated as DLinear algorithm.

[0074] 2.2 Simulation Experiment 2 is a simulation of the Weather dataset.

[0075] Simulation Experiment 2 of this invention uses the method of this invention and three existing technologies to simulate the Weather dataset, obtaining time series prediction results for historical input sequence lengths of 336 and predicted future sequence lengths of 96, 192, 336, and 720, respectively. These results are then compared with real data to obtain the MSE (Mean Sequence Estimate). The obtained scores are then plotted as shown in the figure. Figure 5 The bar chart shown.

[0076] The three existing technologies used in Simulation Experiment 2 are: DLinear algorithm, PatchTST algorithm, and TimesNet algorithm, with the DLinear algorithm being the same as in Simulation Experiment 1. The other two existing technologies used are:

[0077] The existing technology, PatchTST algorithm, is a long-term forecasting method proposed by Nie Y et al. in their paper "A time series is worth 64 words: Long-term forecasting with transformers" (International Conference on Learning Representations (ICLR) 2023), abbreviated as PatchTST algorithm.

[0078] The existing technology, the TimesNet algorithm, is a time series analysis method proposed by Wu H et al. in their paper "TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis" (International Conference on Learning Representations (ICLR), 2023), and is referred to as the TimesNet algorithm for short.

[0079] 2.3 Simulation Experiment 3 is a simulation of the Electricity power consumption dataset.

[0080] Simulation Experiment 3 of this invention uses the method of this invention and four existing technologies: DLinear algorithm, PatchTST algorithm, TimesNet algorithm, and Informer algorithm to simulate the Electricity electricity consumption dataset. It obtains time series prediction results for historical input sequence lengths of 336 and predicted future sequence lengths of 96, 192, 336, and 720, respectively. These results are compared with real data to obtain the MSE (Mean Sequence Estimate). The obtained scores are then plotted as shown in the figure. Figure 6 The bar chart shown.

[0081] In simulation experiment 3, the four existing technologies used are the same as those mentioned in simulation experiments 1 and 2.

[0082] The effects of the present invention will be further described below with reference to the provided simulation diagrams.

[0083] Figure 4 The image shows the MSE results on the Traffic dataset. The horizontal axis represents the predicted future length, and the vertical axis represents the mean squared error (MSE). The individual bar charts illustrate the relationship between the MSE obtained using the method proposed in this invention and the predicted future length, as well as the relationship between the MSE obtained using existing technologies DLinear and Informer simulations and the predicted future length. Figure 4 As can be seen, the MSE obtained by the method of the present invention is always lower than that of the simulation results of the prior art DLinear and Informer for all predicted future lengths, especially when the long-term prediction (such as 720 time steps) is more obvious.

[0084] Figure 5 The MSE results on the Weather dataset are presented. The physical meanings of the x and y axes are the same as before. The individual bar charts represent the relationship between the mean squared error (MSE) obtained from simulations using this invention and existing technologies such as PatchTST, DLinear, and TimesNet, and the predicted future length. Figure 5 The results show that the MSE value of the method of the present invention is superior to other existing technologies across all prediction time intervals. In particular, the method of the present invention exhibits lower errors compared to PatchTST and TimesNet in long-term predictions (336 and 720 time steps), thereby highlighting the ability of the method of the present invention to maintain prediction accuracy over a long period of time.

[0085] Figure 6 The table shows the MSE results on the Electricity dataset. The physical meanings of the x and y axes are the same as before. The individual bar charts represent the relationship between the mean squared error (MSE) and the predicted future length obtained from simulations using this invention and existing techniques such as PatchTST, DLinear, TimesNet, and Informer. Figure 6 It can be seen that although all methods perform similarly in shorter prediction times (96 and 192 time steps), the method of this invention shows a significantly lower MSE in longer prediction intervals (336 and 720 time steps), demonstrating its efficiency and accuracy in handling predictions of complex dynamic systems.

[0086] In summary, the simulation results of this invention demonstrate the effectiveness and superiority of the proposed method on various time series datasets, particularly in long-term forecasting tasks, and its applicability to multiple key fields such as power, transportation, and weather. Furthermore, this invention optimizes the model's generalization ability and adaptability through bidirectional normalization and channel independence design, enabling it to achieve consistently high performance in complex dynamic system application scenarios.

Claims

1. A long-term series forecasting method based on channel independence and bidirectional normalization, applied to power load forecasting, traffic flow forecasting, and weather forecasting, characterized in that... The prediction method involves constructing a bidirectional normalized learnable subnetwork consisting of a forward normalization learning module and a backward normalization learning module; and constructing a channel-independent dual-path prediction subnetwork. The steps of this prediction method are as follows: Step 1: Build two multilayer perceptron modules with identical structures. Set one multilayer perceptron module as a forward normalization learning module and the other multilayer perceptron module as a reverse normalization learning module to obtain a bidirectional normalized learnable subnetwork. Step 2: Construct a channel-independent dual-path prediction subnetwork by connecting the multilayer perceptron prediction module and the sparse attention prediction module in parallel, and then connecting them in series with the dual-path output fusion module. Step 3: A long-term series prediction network is formed by sequentially connecting the forward normalization learning module, the channel-independent dual-path prediction sub-network, and the inverse normalization learning module. Step 4, construct a multivariate time series dataset: The sliding window segmentation method is used to segment historical data for predicting future time series. A multivariate time series dataset was obtained. The dataset is divided into training, validation, and test sets in chronological order. This represents the total number of samples in a multivariate time series dataset. Indicates the first Multidimensional historical data at any given moment Indicates the length of the historical data sequence. Dimensions representing historical data; Step 5: Adaptively adjust network parameters: Set up the training environment for the long-term series prediction network, input the training set into the long-term series prediction network for computation, use the gradient function to iteratively update the network parameters, and fine-tune the model using the validation set until the network loss function converges. Evaluate the model's performance metrics using the test set to obtain the trained time series prediction network. Step 6: Using the same method as in Step 4, the data of the long-term series to be predicted is segmented by sliding window and then input into the trained time series prediction network to output the prediction result.

2. The long-term series prediction method based on channel independence and bidirectional normalization according to claim 1, characterized in that, The forward normalization learning module described in step 1 consists of a first linear layer, an activation layer, and a second linear layer connected in series; the input dimension of the first linear layer is... The output is set to 64, the activation layer uses the ReLU function, the input dimension of the second linear layer is set to 64, the output size is 2, and the parameters and bias of the second linear layer are initialized to 0. The value of is equal to the length of the historical time series of a sample.

3. The long-term series prediction method based on channel independence and bidirectional normalization according to claim 1, characterized in that, The inverse normalization learning module described in step 1 consists of a first linear layer, an activation layer, and a second linear layer connected in series; the input dimension of the first linear layer is... The output dimension is set to 64, the activation layer uses the ReLU function, the input dimension of the second linear layer is set to 64, the output dimension is 2, and the parameters and bias of the second linear layer are initialized to 0. The value of is equal to the length of the future time series to be predicted.

4. The long-term series prediction method based on channel independence and bidirectional normalization according to claim 1, characterized in that, The multilayer perceptron prediction module described in step 2 consists of parallel linear layers of equal number and dimension to the historical data. The input dimension of each linear layer is equal to the length of the historical time series of the sample. The values ​​are equal, and the output dimension is set to the length of the predicted future time series. .

5. The long-term series prediction method based on channel independence and bidirectional normalization according to claim 4, characterized in that, The sparse attention prediction module described in step 2 has the following structure: a linear transformation layer, an encoder attention layer, and a linear projection layer. The input dimension of the linear transformation layer of the sparse attention prediction module is compared with the length of the historical time series. The values ​​of are equal, the output dimension is set to 512, the attention layer scaling factor dropout is set to 0.3, the hidden layer dimension is set to 2048, the activation function is implemented using the GELU function, the input dimension of the linear projection layer is set to 512, and the output dimension is equal to the length of the predicted future time series. The values ​​are equal.

6. The long-term series prediction method based on channel independence and bidirectional normalization according to claim 5, characterized in that, The structure of the dual-output fusion module mentioned in step 2 is as follows: There are learnable parameters, among which... To match the dimensional values ​​of historical data, the parameters of the learnable parameter layer of the dual-output fusion module are all set to 0.

5.

7. The long-time series prediction method based on channel independence and bidirectional normalization according to claim 1, characterized in that, The ratio of the training set, validation set, and test set in step 4 is 8:1:

1.

8. The long-term series prediction method based on channel independence and bidirectional normalization according to claim 1, characterized in that, The calculation steps of the long-term series prediction network described in step 5 are as follows: The first step is to process the original multidimensional time series samples of the prediction task. The input is fed into a long-term series prediction network. A forward normalization learning module normalizes the original multidimensional time series to eliminate distribution shifts between the original input sequences, resulting in a multidimensional time series with input shift repair. The expression is as follows: ； in, This indicates that the input is time series data. and These represent the learnable scaling and offset parameters, respectively. The symbol for element-wise multiplication. and These represent the mean and standard deviation, respectively. The second step is to input the offset-corrected multidimensional time series data. The input is a channel-independent dual-path prediction subnetwork, with one path inputting a channel-independent multilayer perceptron prediction module. Each channel is then fed into its corresponding multilayer perceptron to extract independent feature representations for each channel, resulting in a channel-independent result matrix. Its expression is as follows: ； ； in, Indicates the first Each channel feature vector Indicates the first Multilayer perceptron computation in multiple dimensions Indicates the total number of channels. The result matrix is ​​channel-independent; The other path in the channel-independent dual-path prediction subnetwork, via the sparse attention network prediction module, captures the correlation and covariate information between channels in each channel sequence, thus obtaining the result of the sparse attention mechanism. That is, the result matrix of multi-channel attention prediction, its expression is as follows: ； in, This represents the result matrix of attention prediction across multiple channels. This indicates the computation of the mapping layer. This represents sparse attention computation. This represents the calculation of the linear transformation layer; Step 3: Channel-independent result matrix The result matrix of inter-channel attention prediction The common input dual-output fusion module obtains the dual-channel fusion result matrix, the expression of which is as follows: ； in, This represents the result matrix of the dual-path fusion. This represents the computation of dual-path fusion; The fourth step involves inverse normalization of the dual-path fusion result matrix via an inverse normalization learning module to restore the original scale of the prediction results and eliminate the distribution offset between the input and output. The expression is as follows: ； in, This represents a multidimensional time series that ultimately predicts the future. and Let represent the learnable mean and standard deviation, respectively. and This represents the learnable scaling and offset parameters.

9. The long-term series prediction method based on channel independence and bidirectional normalization according to claim 1, characterized in that, The training environment for the long-term series prediction network described in step 5 is as follows: the network loss function is the mean squared error (MSE), the optimizer is Adam, the initial learning rate is set to 0.001, the batch size is set to 32, and the EarlyStopping training process is set to 10 epochs.

10. The long-term series prediction method based on channel independence and bidirectional normalization according to claim 1, characterized in that, The network loss function described in step 5 is as follows: ； in, This represents the true future sequence of the training set samples. This represents the future sequence ultimately predicted by the long-term time-series forecasting model. This represents the total number of channels that have the same dimension value as the historical data. Indicates the length of the predicted future sequence. Indicates the channel number. Indicates the sequence number of a future time point.

Images