Time sequence prediction method based on FFT and interactive depth separable convolution

Abstract

The invention belongs to the technical field of time sequence prediction, and particularly relates to a time sequence prediction method based on FFT and interactive depth separable convolution. Frequency domain features and time domain features are separated, so that the model can model two types of features respectively, and the prediction stability and accuracy are improved; by introducing fast Fourier transform and a random frequency mask mechanism, the capturing capability of the model on frequency domain features is effectively improved; by designing an interactive depth separable convolution block, multi-scale time sequence features can be efficiently extracted, and meanwhile, the calculation complexity is effectively reduced. The original features are reserved through residual connection, information degradation is prevented, and the combination of time domain and frequency domain features is more flexible by adapting to a fusion strategy.

Description

Technical Field

[0001] The present invention belongs to the technical field of time series prediction, and in particular relates to a time series prediction method based on FFT (Fast Fourier Transform) and interactive depth separable convolution. Background Art

[0002] Time series forecasting has attracted much attention due to its wide application in areas such as traffic flow. Accurate time series forecasting can help decision makers identify trends in advance, optimize resource allocation, and reduce potential risks. However, traditional methods often have difficulty effectively capturing features in the time and frequency domains when processing complex high-dimensional time series data, and face many challenges.

[0003] Traditional time series forecasting methods are mainly based on statistical models: AR (autoregressive model) is a common linear model in the field of time series analysis, which uses observations from past time points to predict future values. MA (moving average model): It assumes that the value of the current time series is the result of a weighted average of a series of previous random error terms (noise). Unlike the autoregressive model (AR), the MA model does not directly use past observations to predict, but uses error terms. However, relying solely on past error terms is powerless for the long-term trend of the data. The lag order needs to be manually selected. It cannot be directly used for non-stationary data. ARIMA (autoregressive difference moving average model): It combines the AR and MA models and introduces difference operations to solve the problem of non-stationary series. Parameters need to be manually selected, and parameter adjustment is more complicated. It cannot capture nonlinear relationships in time series. The computational complexity is high and the training time is long. The modeling ability for long-term dependence and high-dimensional data is limited.

[0004] Deep learning-based methods, such as recurrent neural networks (RNNs) and long short-term memory (LSTMs), have demonstrated strong capabilities in capturing temporal dependencies. However, these models have some limitations, such as slow training speed, difficulty in capturing long-range dependencies, and low efficiency when processing large-scale data. Recent studies, such as Transformer and spectrum-based methods, have made significant progress through parallel computing and frequency domain feature extraction. However, how to efficiently fuse temporal features with spectral domain enhancement information remains a difficult problem in current research. Summary of the invention

[0005] In order to overcome the problems in the prior art, the present invention proposes a time series prediction method based on FFT and interactive depth-separable convolution.

[0006] The technical solution of the present invention to solve the above technical problems is as follows: In a first aspect, the present invention provides a time series prediction method based on FFT and interactive depth separable convolution, comprising the following steps: Preprocess the historical traffic flow observation time series of the input urban intersection to obtain the preprocessed traffic flow observation time series; Construct a traffic flow time series prediction model. The processing steps in the constructed traffic flow time series prediction model include: performing layer normalization and linear processing on the preprocessed traffic flow observation time series to obtain the linearly transformed traffic flow observation time series; performing frequency domain processing and feature extraction on the linearly transformed traffic flow observation time series and converting it back to the time domain to obtain a time domain signal; extracting interaction features of the time domain signal through an interactive depthwise separable convolution block and performing mapping to obtain a mapped output feature; Perform residual connection on the preprocessed traffic flow observation time series and the mapped output feature, and calculate to obtain a residual fusion feature; convert the residual fusion feature into a prediction value through a fully connected layer and perform denormalization processing to obtain the final prediction result.

[0007] Further, the preprocessing of the historical traffic flow observation time series of the input urban intersection to obtain the preprocessed traffic flow observation time series includes: Remove the mean from the historical traffic flow observation time series of the input urban intersection to obtain the historical traffic flow observation time series after mean removal; Normalize the historical traffic flow observation time series after mean removal, and further obtain the preprocessed traffic flow observation time series.

[0008] Further, the layer normalization and linear processing of the preprocessed traffic flow observation time series to obtain the linearly transformed traffic flow observation time series includes: For the preprocessed traffic flow observation time series, calculate the mean and standard deviation: Based on the calculated mean and standard deviation, perform the first normalization on the preprocessed traffic flow observation time series to obtain the first normalized traffic flow observation time series; Perform a linear transformation on the first normalized traffic flow observation time series to obtain the linearly transformed traffic flow observation time series.

[0009] Further, the frequency domain processing and feature extraction of the linearly transformed traffic flow observation time series and converting it back to the time domain to obtain a time domain signal includes: Use the fast Fourier transform to transform the linearly transformed traffic flow observation time series to the frequency domain to obtain the traffic flow observation time series after fast Fourier transform; Random frequency masking, generating a random binary masking matrix; The traffic flow observation time series after fast Fourier transformation is multiplied by the binary masking matrix element by element according to the principal element to obtain the masking data; Construct learnable global complex weights and learnable local complex weights; The frequency domain representation of the linearly transformed traffic flow observation time series is weighted by a learnable global complex weight to obtain the complex weighted frequency domain traffic flow feature: The mask data is multiplied by the learnable local complex weight to obtain the local complex mask feature; The complex weighted frequency domain traffic flow features and local complex masking features are integrated to capture comprehensive frequency details and obtain the complex fused frequency domain-masked traffic flow features: The complex fused frequency domain-masked traffic flow features are converted back to the time domain and the inverse fast Fourier transform is applied to obtain the time domain signal.

[0010] Furthermore, the interactive features of the time domain signal are extracted and mapped by the interactive depth-separable convolution block to obtain the mapped output features, including: The time domain signal is used as the input of an interactive depth-wise separable convolution block, wherein the interactive depth-wise separable convolution block includes depth-wise separable convolution, point-wise convolution, activation and Dropout; the interactive depth-wise separable convolution block includes a first group of depth-wise separable convolutions and a second group of depth-wise separable convolutions; The output of the first set of point convolutions is multiplied by the output of the second set of activations and Dropout to obtain the first interaction feature: The second set of point convolutions is multiplied by the output of the first set of activations and Dropout to obtain the second interaction feature: The first interaction feature and the second interaction feature are added together as the final interaction feature: The final interaction features are mapped back to the input feature dimension through ordinary convolution to obtain the mapped output features.

[0011] Furthermore, the residual fusion feature is converted into a prediction value through a fully connected layer, and denormalized to obtain a final prediction result, including: The residual fusion features are converted into prediction values ​​through a fully connected layer: ; in, represents the weight matrix; Represents the bias vector, which is used to adjust the output value; represents the output of the linear layer; Represents residual fusion features; Denormalize the output of the linear layer to get the final prediction result: ; in, Indicates the final prediction result; represents the mean of the time series; Represents the standard deviation of the time series.

[0012] Furthermore, it also includes: training the traffic flow time series prediction model to obtain a trained traffic flow time series prediction model; inputting the time series data to be predicted into the trained traffic flow time series prediction model to obtain a prediction result.

[0013] Compared with the prior art, the present invention has the following technical effects: (1) The present invention separates frequency domain features from time domain features, so that the model can model the two types of features separately, thereby improving the stability and accuracy of prediction; (2) By introducing the fast Fourier transform and random frequency masking mechanism, the model's ability to capture frequency domain features is effectively improved. This module can randomly discard frequency components during training, forcing the model to learn more robust features from multiple frequency bands and improve its ability to resist noise and outliers. Through this strategy, the model can not only effectively capture high-frequency and low-frequency information, but also improve its performance on unseen data.

[0014] (3) By designing an interactive deep separable convolution block, it is possible to efficiently extract multi-scale temporal features while effectively reducing computational complexity. Compared with traditional convolutional networks, deep separable convolution significantly reduces the number of parameters and computational cost by decomposing the convolution operation into two parts: deep convolution and point-by-point convolution, making it suitable for resource-constrained application scenarios.

[0015] (4) The present invention retains the original features through residual connection to prevent information degradation, and at the same time adapts the fusion strategy to make the combination of time domain and frequency domain features more flexible. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are 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 flowchart of a time series prediction method based on FFT and interactive depth separable convolution of the present invention; Figure 2 It is the predicted fitting diagram of the traffic flow curve; Figure 3 It is a structural schematic diagram of the time series prediction method based on FFT and interactive deep separable convolution of the present invention. DETAILED DESCRIPTION

[0018] In order to further explain the technical means and effects taken by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. The specific features, structures or characteristics in one or more embodiments may be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present invention.

[0019] In one embodiment of the present invention, referring to Figure 1-Figure 3 , a time series prediction method based on FFT and interactive depthwise separable convolution is provided, comprising the following steps: Preprocess the input historical traffic flow observation time series of the urban intersection to obtain the preprocessed traffic flow observation time series; A traffic flow time series prediction model is constructed. The processing steps in the constructed traffic flow time series prediction model include: performing layer normalization and linear processing on the preprocessed traffic flow observation time series to obtain the traffic flow observation time series after linear transformation; performing frequency domain processing and feature extraction on the traffic flow observation time series after linear transformation, and converting it back to the time domain to obtain the time domain signal; extracting the interactive features of the time domain signal through an interactive deep separable convolution block and mapping it to obtain the mapping output feature; performing residual connection on the preprocessed traffic flow observation time series and the mapping output feature to calculate the residual fusion feature; converting the residual fusion feature into a predicted value through a fully connected layer, and performing denormalization processing to obtain the final prediction result; Training the traffic flow time series prediction model to obtain a trained traffic flow time series prediction model; The time series data to be predicted is input into the trained traffic flow time series prediction model to obtain a prediction result.

[0020] The following is a detailed explanation of each of the above steps: Step 100: preprocessing the input historical traffic flow observation time series of the urban intersection to obtain the preprocessed traffic flow observation time series.

[0021] The preprocessing includes mean removal and standardization, and uses a sliding window method to slice the input historical traffic flow observation time series of the urban intersection to obtain a plurality of subsequences of fixed length.

[0022] In this embodiment, a specific implementation of step 100 may be: Step 110: Perform mean removal and standardization on the input historical traffic flow observation time series of the urban intersection to ensure stable data distribution and improve the model convergence speed.

[0023] The original time series is , the calculation formula for mean removal is as follows: ; in, represents the mean of the time series; Represents the historical traffic flow observation time after mean removal.

[0024] Standardize the historical traffic flow observation time series after mean removal: ; in, represents the standardized traffic flow observation time; represents the standard deviation of the time series, , T represents the total length of the time series data.

[0025] Based on the standardized traffic flow observation time, the preprocessed traffic flow observation time series is obtained. .

[0026] Step 120: Perform overlapping block processing on the standardized traffic flow observation time series to obtain a number of subsequences of fixed length.

[0027] Specifically, subsequences of length w are extracted from the standardized traffic flow observation time series in a sliding window manner, and each window partially overlaps with the previous window to obtain several subsequences of fixed length.

[0028] The standardized traffic flow observation time series is divided into blocks according to the fixed window length w and step length s, and the subsequences centered on the time step t are extracted. The range of each subsequence is ,in, , T is the total length of the time series data.

[0029] Through overlapping blocks, each subsequence contains some of the same data points, so that the correlation between adjacent time windows is preserved and key time information is not lost due to division.

[0030] Through the sliding window method, each time point may be used by multiple subsequences, increasing the number of effective training samples and improving the generalization ability of the model. Alleviate data sparsity: When there is less data in certain time periods, overlapping blocks can ensure that each time point is not ignored, thereby improving data coverage.

[0031] The sliding window method can ensure that the local features of the time series will not be truncated due to window division, which helps the model learn short-term and long-term dependencies.

[0032] Since multiple windows share some of the same data points, the prediction results are integrated during the calculation process of multiple windows, thereby reducing the impact of sudden changes on the prediction.

[0033] Step 200: Construct a traffic flow time series prediction model.

[0034] Construct a traffic flow time series prediction model, refer to Figure 3 The traffic flow time series prediction model includes a first-layer normalization module, a frequency domain enhancement module, a second-layer normalization module, an interactive depth separable convolution block, a residual module and a fully connected layer; the input of the first-layer normalization module receives the preprocessed traffic flow observation time series, the output of the first-layer normalization module is connected to the input of the first-layer normalization module, the output of the first-layer normalization module is connected to the input of the frequency domain enhancement module, the output of the frequency domain enhancement module is connected to the input of the second-layer normalization module, the output of the second-layer normalization module is connected to the input of the interactive depth separable convolution block, the output of the interactive depth separable convolution block is connected to the first input of the residual module, the second input of the residual module receives the preprocessed traffic flow observation time series, the output of the residual module is connected to the input of the fully connected layer, and the output of the fully connected layer is used to output the result.

[0035] The first layer normalization module is used to perform layer normalization on the preprocessed traffic flow observation time series; the frequency domain enhancement module is used to perform frequency domain enhancement processing; the second layer normalization module is used to continue layer normalization; the interactive deep separable convolution block is used for time series feature extraction; the residual module is used to perform residual connection on the preprocessed traffic flow observation time series and the interactive deep separable convolution, and calculate the fused features; the fully connected layer is used to convert the fused features into predicted values, and perform denormalization processing to obtain the final prediction result.

[0036] As a specific example, this step 200 may include the following sub-steps: Step 210: performing layer normalization and linear processing on the preprocessed traffic flow observation time series to obtain a linearly transformed traffic flow observation time series.

[0037] Layer normalization normalizes all features of each preprocessed traffic flow observation time series sample instead of normalizing batch data. This helps stabilize the network training process and is suitable for time series data with large batch size variations.

[0038] In this embodiment, a specific implementation of step 210 may be: Step 2101: For the preprocessed traffic flow observation time series (shape is (B, T, D), D is the feature dimension), layer normalization calculates the mean and standard deviation on the last dimension D: ; ; In the above formula, represents the mean of the preprocessed traffic flow observation time series; Represents the standard deviation of the preprocessed traffic flow observation time series.

[0039] Step 2102: Perform the first normalization on the preprocessed traffic flow observation time series to obtain the first normalized traffic flow observation time series: ; in, represents the first normalized traffic flow observation time; is a very small number used to prevent division by zero errors.

[0040] Step 2103: Perform a linear transformation on the first normalized traffic flow observation time series: ; in, represents the traffic flow observation time after linear transformation, represents the scaling parameter, Indicates the offset parameter.

[0041] Traffic flow observation time after linear transformation Then we get the traffic flow observation time series after linear transformation .

[0042] Step 220: Perform frequency domain processing and feature extraction on the linearly transformed traffic flow observation time series, and convert it back to the time domain to maintain the consistency of the data structure.

[0043] The linearly transformed traffic flow observation time series is converted to the frequency domain using the fast Fourier transform. Two learnable parameter matrices are constructed, and the adaptive weight matrix is ​​generated through the parameter matrix; the adaptive weight matrix is ​​used to filter the frequency domain signal, emphasizing the enhancement or suppression of certain frequency components, helping the network to pay more attention to important frequency components. The adaptive weight matrix is ​​used to filter the frequency domain signal, and the purpose of filtering is to emphasize the feature extraction of high-frequency or low-frequency parts.

[0044] At the same time, a random masking strategy is used to adaptively adjust the frequency domain components to enhance the robustness and diversity of the model. Through the masking operation, different frequency components can be randomly selected, making the model more flexible. After completing the frequency domain enhancement, the frequency domain features after filtering and random masking are fused to obtain a more comprehensive feature representation.

[0045] In this embodiment, a specific implementation of step 220 may be: Step 2201: Use fast Fourier transform to transform the traffic flow observation time series after linear transformation into frequency domain, and obtain the traffic flow observation time series after fast Fourier transform: ; In the above formula, represents the traffic flow observation time series after linear transformation; represents the traffic flow observation time series after fast Fourier transformation, Indicates the number of frequency points; F is the traffic flow observation time series after linear transformation Frequency domain representation of .

[0046] Step 2202: Use a random masking strategy to adaptively adjust the traffic flow observation time series after fast Fourier transformation to enhance the robustness and diversity of the model.

[0047] Random frequency masking generates a random binary masking matrix M, the shape of which is the same as the linearly transformed traffic flow observation time series. The frequency domain representation is consistent, and the value of each element is 0 or 1, indicating whether the frequency component is retained.

[0048] ; in, It means that each frequency component has a 50% probability of being retained; Represents a Bernoulli distribution that generates a binary number (0 or 1).

[0049] The traffic flow observation time series after fast Fourier transformation is multiplied element by element with the binary masking matrix M according to the main element to obtain the masking data: ; In the above formula, Represents masked data; ⊙ represents element-by-element multiplication of the main elements.

[0050] Step 2203: construct two learnable parameter matrices, and generate an adaptive weight matrix through the parameter matrix; the adaptive weight matrix is ​​used to filter the frequency domain signal, emphasizing the enhancement or suppression of certain frequency components, and helping the network to pay more attention to important frequency components. The adaptive weight matrix is ​​used to filter the frequency domain signal, and the purpose of filtering is to emphasize the feature extraction of high-frequency or low-frequency parts.

[0051] The two learnable parameter matrices include learnable global complex weights and learnable local complex weights. is a learnable global complex weight, represents learnable local complex weights; Through learnable global complex weights The traffic flow observation time series after linear transformation Frequency domain representation of F After weighting, we get the complex weighted frequency domain traffic flow characteristics: ; Masking Data With learnable local complex weights Multiply them together to get the local complex masking feature: ; In the above formula, represents the complex weighted frequency domain traffic flow characteristics; Represents a local complex masking feature.

[0052] Step 2204: Integrate the complex weighted frequency domain traffic flow features and the local complex masking features to capture comprehensive frequency details and obtain the complex fused frequency domain-masked traffic flow features: ; In the above formula, Represents the complex fused frequency domain-masked traffic flow characteristics.

[0053] Step 2205: Convert the complex fused frequency domain-masked traffic flow features back to the time domain and apply the inverse fast Fourier transform (IFFT) to obtain the time domain signal : ; IFFT ensures that the enhanced features remain consistent with the original data structure of the input time series.

[0054] Step 230: performing layer normalization on the time domain signal after the frequency domain enhancement processing.

[0055] The layer normalization after frequency domain enhancement is used to standardize the features after frequency domain enhancement to ensure the stability of the model training process.

[0056] Step 240: extract the interactive features of the time domain signal through an interactive depth-wise separable convolution block and map them to obtain mapped output features.

[0057] Combining convolution operations with different receptive fields can extract features from different time scales. After extracting features at multiple scales, the feature interaction mechanism is used to fuse features at different scales. By calculating the interaction between features at different scales, more complex temporal dependencies can be captured.

[0058] In this embodiment, a specific implementation of step 240 may be: Step 2401: The time domain signal S is used as the input of an interactive depthwise separable convolution block, wherein the interactive depthwise separable convolution block includes a first set of depthwise separable convolutions and a second set of depthwise separable convolutions.

[0059] The first set of depthwise separable convolutions: The first set of depthwise convolutions: ; The first set of point convolutions: ; The first set of activations and Dropout: ; The second set of depthwise separable convolutions: The second set of depthwise convolutions: ; The second set of point convolutions: ; The second set of activations and Dropout: ; In the above formula, ReLU activation function. represents the depthwise convolution, represents point convolution; Represents the first set of depthwise convolution outputs; represents the first set of point convolution; Represents the first set of activation and Dropout output; Represents the second set of depthwise convolution outputs; represents the second set of point convolution; Represents the second set of activations and Dropout outputs.

[0060] Among them, the first group of deep convolutions and the second group of deep convolutions use two groups of deep convolutions with different kernel sizes (3 and 5) to extract features in different receptive field ranges, and the point convolution uses one-dimensional convolution to mix information across channels.

[0061] Step 2402: feature interaction, feature interaction is achieved through cross-point multiplication to obtain two interactive features.

[0062] The output of the first set of point convolutions is multiplied by the output of the second set of activations and Dropout to obtain the first interaction feature: ; in, Represents the first interaction feature.

[0063] The second set of point convolutions is multiplied by the first set of activations and the Dropout output to obtain the second interaction feature: ; in, Represents the second interaction feature.

[0064] Add the first interaction feature and the second interaction feature as the final interaction feature: ; in, represents the final interaction feature.

[0065] Step 2403: Map the final interaction features back to the input feature dimension through ordinary convolution to obtain the mapped output features: ; in, The convolution kernel size is 1×1; Represents the mapping output features.

[0066] Step 250: further extract global time series features through a multi-layer neural network, and fuse the features extracted from each layer. This step aims to integrate information at different levels to obtain global time series information; The features of the previous layer mapping output are further extracted, and the global time series features are extracted through a multi-layer neural network. The features extracted from each layer are fused to integrate information at different levels, thereby obtaining complete global time series information.

[0067] Step 260: Perform residual connection on the preprocessed traffic flow observation time series and the mapped output features of the interactive deep separable convolution output to calculate the residual fusion features.

[0068] To ensure the integrity of the model and information is not lost during feature processing, a residual connection mechanism is used to convert the preprocessed traffic flow observation time series into Mapping output features with the output of an interactive deep separable convolutional network Addition: ; In the above formula, Represents the residual fusion feature.

[0069] Residual connections are crucial in the deep structure of the model. They can alleviate the problem of information degradation in multi-layer propagation, while retaining the key information of the original features and improving the learning stability of the model.

[0070] Step 270: The residual fusion features are converted into prediction values ​​through a fully connected layer, and denormalized to obtain the final prediction result.

[0071] In this embodiment, a specific implementation of step 270 may be: Step 2701: The residual fusion features are converted into prediction values ​​through a fully connected layer: ; in, Represents the weight matrix, which determines the mapping of input features to output features; Represents the bias vector, which is used to adjust the output value; represents the output of the linear layer.

[0072] Step 2702: Denormalize to obtain the final prediction result: ; In the above formula, Represents the final prediction result, which combines the characteristics of frequency domain and time domain, and can achieve excellent results in prediction accuracy and data pattern preservation.

[0073] Step 300: training the traffic flow time series prediction model to obtain a trained traffic flow time series prediction model.

[0074] Step 400: input the time series data to be predicted into the trained traffic flow time series prediction model to obtain a prediction result.

[0075] Reference Figure 2 , the blue curve represents the true value, and the orange curve represents the predicted value of the model. Overall, the predicted value is highly consistent with the true value, indicating that the present invention can also ensure accuracy and effective fitting effect when predicting long-sequence traffic flow.

[0076] 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, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such 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 be included in the protection scope of the present application.

Claims

1. A time series prediction method based on FFT and interactive depth-separable convolution, characterized in that: The following steps are involved: Preprocess the input historical traffic flow observation time series of the urban intersection to obtain the preprocessed traffic flow observation time series; A traffic flow time series prediction model is constructed. The processing steps in the constructed traffic flow time series prediction model include: performing layer normalization and linear processing on the preprocessed traffic flow observation time series to obtain the traffic flow observation time series after linear transformation; performing frequency domain processing and feature extraction on the traffic flow observation time series after linear transformation, and converting it back to the time domain to obtain a time domain signal; extracting the interactive features of the time domain signal through an interactive deep separable convolution block and mapping it to obtain a mapping output feature; The preprocessed traffic flow observation time series and the mapped output features are residually connected to calculate the residual fusion features; the residual fusion features are converted into prediction values ​​through a fully connected layer and denormalized to obtain the final prediction results.

2. According to claim 1, a time series prediction method based on FFT and interactive depth-separable convolution is characterized in that: The preprocessing of the input historical traffic flow observation time series of the urban intersection to obtain the preprocessed traffic flow observation time series includes: Perform mean removal on the input historical traffic flow observation time series of the urban intersection to obtain the historical traffic flow observation time series after mean removal; The historical traffic flow observation time series after mean removal is standardized to obtain the preprocessed traffic flow observation time series.

3. The time series prediction method based on FFT and interactive depth-separable convolution according to claim 1, characterized in that: The layer normalization and linear processing are performed on the preprocessed traffic flow observation time series to obtain the traffic flow observation time series after linear transformation, including: For the preprocessed traffic flow observation time series, calculate the mean and standard deviation: Based on the calculated mean and standard deviation, the preprocessed traffic flow observation time series is normalized for the first time to obtain the first normalized traffic flow observation time series; The first normalized traffic flow observation time series is linearly transformed to obtain the traffic flow observation time series after linear transformation.

4. The time series prediction method based on FFT and interactive depth-separable convolution according to claim 3, characterized in that: The traffic flow observation time series after linear transformation is processed in the frequency domain and features are extracted, and converted back to the time domain to obtain a time domain signal, including: The traffic flow observation time series after linear transformation is converted into the frequency domain using fast Fourier transform to obtain the traffic flow observation time series after fast Fourier transform; Random frequency masking, generates a random binary masking matrix; The traffic flow observation time series after fast Fourier transformation is multiplied by the binary masking matrix element by element according to the principal element to obtain the masking data; Construct learnable global complex weights and learnable local complex weights; The frequency domain representation of the linearly transformed traffic flow observation time series is weighted by a learnable global complex weight to obtain the complex weighted frequency domain traffic flow feature: The mask data is multiplied by the learnable local complex weight to obtain the local complex mask feature; The complex weighted frequency domain traffic flow features and local complex masking features are integrated to capture comprehensive frequency details and obtain the complex fused frequency domain-masked traffic flow features: The complex fused frequency domain-masked traffic flow features are converted back to the time domain and the inverse fast Fourier transform is applied to obtain the time domain signal.

5. The time series prediction method based on FFT and interactive depth-separable convolution according to claim 4, characterized in that: The interactive features of the time domain signal are extracted and mapped by the interactive depth-separable convolution block to obtain the mapped output features, including: The time domain signal is used as the input of an interactive depth-wise separable convolution block, wherein the interactive depth-wise separable convolution block includes depth-wise separable convolution, point-wise convolution, activation and Dropout; the interactive depth-wise separable convolution block includes a first group of depth-wise separable convolutions and a second group of depth-wise separable convolutions; The output of the first set of point convolutions is multiplied by the output of the second set of activations and Dropout to obtain the first interaction feature: The second set of point convolutions is multiplied by the output of the first set of activations and Dropout to obtain the second interaction feature: The first interaction feature and the second interaction feature are added together as the final interaction feature: The final interaction features are mapped back to the input feature dimension through ordinary convolution to obtain the mapped output features.

6. The time series prediction method based on FFT and interactive depth-separable convolution according to claim 5, characterized in that: The residual fusion features are converted into prediction values ​​through a fully connected layer, and denormalized to obtain the final prediction results, including: The residual fusion features are converted into prediction values ​​through a fully connected layer: ； in, represents the weight matrix; Represents the bias vector, which is used to adjust the output value; represents the output of the linear layer; Represents residual fusion features; Denormalize the output of the linear layer to get the final prediction result: ； in, Indicates the final prediction result; represents the mean of the time series; Represents the standard deviation of the time series.

7. The time series prediction method based on FFT and interactive depth-separable convolution according to claim 1, characterized in that: The method further includes: training the traffic flow time series prediction model to obtain a trained traffic flow time series prediction model; inputting the time series data to be predicted into the trained traffic flow time series prediction model to obtain a prediction result.

Images