Time Series Prediction Method Based on FFT and Interactive Depthwise Separable Convolution

Abstract

The present invention belongs to the technical field of time series prediction, and particularly relates to a time series prediction method based on FFT and interactive depthwise separable convolution. By separating the frequency domain features from the time domain features, the model can model the two types of features separately, improving the stability and accuracy of prediction; by introducing the fast Fourier transform and the random frequency masking mechanism, the model's ability to capture frequency domain features is effectively enhanced; by designing the interactive depthwise separable convolution block, multi-scale time series features can be efficiently extracted while effectively reducing the computational complexity. The original features are retained through residual connections to prevent information degradation, and at the same time, the adaptation fusion strategy makes the combination of time domain and frequency domain features more flexible.

Description

Technical Field

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

[0002] Time series prediction has attracted much attention due to its wide applications in fields such as traffic flow. Accurate time series prediction can help decision-makers identify trends in advance, optimize resource allocation, and reduce potential risks. However, traditional methods often face challenges in effectively capturing features in the time domain and frequency domain when dealing with complex high-dimensional time series data.

[0003] Traditional time series prediction methods are mainly based on statistical models:

[0004] AR (Autoregressive Model) predicts future values by using the observed values at past time points and is a common linear model in the field of time series analysis. MA (Moving Average Model): assumes that the value of the current time series is the weighted average of a series of previous random error terms (noise). Different from the autoregressive model (AR), the MA model does not directly use past observed values for prediction but uses error terms. However, relying only on past error terms is powerless for the long-term trend of data. The lag order needs to be manually selected. It cannot be directly used for non-stationary data. ARIMA (Autoregressive Integrated Moving Average Model): combines AR and MA models and introduces differencing operations to solve the non-stationary sequence problem. Parameters need to be manually selected, and parameter tuning is relatively complex. It cannot capture non-linear relationships in time series. The computational complexity is relatively high, and the training time is long. Its ability to model long-term dependencies and high-dimensional data is limited.

[0005] Deep learning-based methods, such as the Recurrent Neural Network (RNN) and Long Short-Term Memory Network (LSTM), have demonstrated powerful capabilities in capturing time dependencies. However, these models have some limitations, such as slow training speed, difficulty in capturing long-range dependencies, and low efficiency in processing large-scale data. Recent research, such as Transformer and spectrum-based methods, has made significant progress through parallel computing and frequency domain feature extraction. However, how to efficiently fuse time domain features with spectrum domain enhanced information remains a difficult problem in current research. Summary of the Invention

[0006] To overcome the problems in the prior art, the present invention proposes a time series prediction method based on FFT and interactive depthwise separable convolution.

[0007] The technical solution of the present invention to solve the above technical problems is as follows:

[0008] In a first aspect, the present invention provides a time series prediction method based on FFT and interactive depthwise separable convolution, comprising the following steps:

[0009] Preprocess the historical traffic flow observation time series of the urban intersection input to obtain a preprocessed traffic flow observation time series;

[0010] 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 a 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;

[0011] Perform residual connection on the preprocessed traffic flow observation time series and the mapped output feature, calculate to obtain a residual fusion feature; convert the residual fusion feature into a predicted value through a fully connected layer and perform denormalization processing to obtain a final prediction result.

[0012] Further, the preprocessing of the historical traffic flow observation time series of the urban intersection input to obtain a preprocessed traffic flow observation time series includes:

[0013] Remove the mean from the historical traffic flow observation time series of the urban intersection input to obtain a historical traffic flow observation time series after mean removal;

[0014] Normalize the historical traffic flow observation time series after mean removal, and further obtain a preprocessed traffic flow observation time series.

[0015] Further, the performing layer normalization and linear processing on the preprocessed traffic flow observation time series to obtain a linearly transformed traffic flow observation time series includes:

[0016] For the preprocessed traffic flow observation time series, calculate the mean and standard deviation:

[0017] Based on the calculated mean and standard deviation, perform the first normalization on the preprocessed traffic flow observation time series to obtain a traffic flow observation time series after the first normalization;

[0018] Perform a linear transformation on the traffic flow observation time series after the first normalization to obtain a linearly transformed traffic flow observation time series.

[0019] Further, the frequency domain processing and feature extraction of the linearly transformed traffic flow observation time series and conversion back to the time domain to obtain a time domain signal include:

[0020] Use the fast Fourier transform to convert the linearly transformed traffic flow observation time series to the frequency domain to obtain the traffic flow observation time series after the fast Fourier transform;

[0021] Random frequency masking to generate a random binary masking matrix;

[0022] Multiply the traffic flow observation time series after the fast Fourier transform element by element with the binary masking matrix according to the main elements to obtain masked data;

[0023] Construct learnable global complex weights and learnable local complex weights;

[0024] Weight the frequency domain representation of the linearly transformed traffic flow observation time series by the learnable global complex weights to obtain complex weighted frequency domain traffic flow features:

[0025] Multiply the masked data by the learnable local complex weights to obtain local complex masked features;

[0026] Integrate the complex weighted frequency domain traffic flow features and the local complex masked features to capture comprehensive frequency details to obtain complex fused frequency domain - masked traffic flow features:

[0027] Convert the complex fused frequency domain - masked traffic flow features back to the time domain by applying the inverse fast Fourier transform to obtain a time domain signal.

[0028] Further, the extraction of interaction features of the time domain signal by the interactive depth - separable convolution block and mapping to obtain a mapped output feature include:

[0029] The time domain signal serves as the input of the interactive depth - separable convolution block, and the interactive depth - separable convolution block includes depth - separable convolution, point convolution, activation, and Dropout; the interactive depth - separable convolution block includes a first group of depth - separable convolution and a second group of depth - separable convolution;

[0030] Multiply the output of the first group of point convolution by the output of the second group of activation and Dropout to obtain a first interaction feature: Multiply the second group of point convolution by the output of the first group of activation and Dropout to obtain a second interaction feature: Add the first interaction feature and the second interaction feature as the final interaction feature:

[0031] Map the final interaction feature back to the input feature dimension through ordinary convolution to obtain a mapped output feature.

[0032] Further, converting the residual fusion feature into a predicted value through a fully connected layer and performing denormalization to obtain the final prediction result includes:

[0033] Converting the residual fusion feature into a predicted value through a fully connected layer:

[0034] ;

[0035] Among them, represents the weight matrix; represents the bias vector used to adjust the output value; represents the output of the linear layer; represents the residual fusion feature;

[0036] Performing denormalization on the output of the linear layer to obtain the final prediction result:

[0037] ;

[0038] Among them, represents the final prediction result; represents the mean of the time series; represents the standard deviation of the time series.

[0039] Further, 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.

[0040] Compared with the prior art, the present invention has the following technical effects:

[0041] (1) The present invention separates the frequency domain features from the time domain features, enabling the model to model the two types of features separately, improving the stability and accuracy of prediction;

[0042] (2) By introducing the fast Fourier transform and the random frequency masking mechanism, the ability of the model to capture frequency domain features is effectively improved. This module can randomly discard frequency components during the training process, forcing the model to learn more robust features from multiple frequency bands, enhancing the anti-interference ability against 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.

[0043] (3) By designing the interactive depthwise separable convolution block, it can efficiently extract multi-scale temporal features while effectively reducing the computational complexity. Compared with traditional convolutional networks, depthwise separable convolution decomposes the convolution operation into depthwise convolution and pointwise convolution, significantly reducing the number of parameters and computational cost, and is suitable for resource-constrained application scenarios.

[0044] (4) The present invention preserves the original features through residual connections 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. Description of the Drawings

[0045] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0046] Figure 1 It is a flowchart of the time series prediction method based on FFT and interactive depthwise separable convolution of the present invention;

[0047] Figure 2 It is a prediction fitting graph of traffic flow curves;

[0048] Figure 3 It is a schematic structural diagram of the time series prediction method based on FFT and interactive depthwise separable convolution of the present invention. Detailed Embodiments

[0049] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in combination with the drawings and preferred embodiments, details the specific embodiments, structures, features, and effects of the technical solutions proposed according to the present invention. The specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0050] In one embodiment of the present invention, referring to Figures 1 - 3 , a time series prediction method based on FFT and interactive depthwise separable convolution is provided, including the following steps:

[0051] Preprocess the historical traffic flow observation time series of the urban intersection input to obtain the preprocessed traffic flow observation time series;

[0052] 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; performing residual connection on the preprocessed traffic flow observation time series and the mapped output feature to calculate a residual fusion feature; converting the residual fusion feature into a prediction value through a fully connected layer and performing denormalization processing to obtain a final prediction result;

[0053] Train the traffic flow time series prediction model to obtain a trained traffic flow time series prediction model;

[0054] Input the time series data to be predicted into the trained traffic flow time series prediction model to obtain a prediction result.

[0055] The following details each of the above steps:

[0056] Step 100: Preprocess the historical traffic flow observation time series of the input urban intersection to obtain the preprocessed traffic flow observation time series.

[0057] The preprocessing includes mean removal and normalization, and slices the historical traffic flow observation time series of the input urban intersection in a sliding window manner to obtain a number of subsequences of a fixed length.

[0058] In this embodiment, a specific implementation manner of step 100 may be:

[0059] Step 110: Perform mean removal and normalization on the historical traffic flow observation time series of the input urban intersection to ensure stable data distribution and improve the model convergence speed.

[0060] The original time series is , and the calculation formula for mean removal is as follows:

[0061] ;

[0062] where represents the mean of the time series; represents the historical traffic flow observation time after mean removal.

[0063] Normalize the historical traffic flow observation time series after mean removal:

[0064] ;

[0065] Among them, represents the standardized traffic flow observation time; represents the standard deviation of the time series, , and T represents the total length of the time series data.

[0066] Based on the standardized traffic flow observation time, the preprocessed traffic flow observation time series is obtained .

[0067] Step 120: Perform overlapping block processing on the standardized traffic flow observation time series to obtain several subsequences of a fixed length.

[0068] Specifically, extract subsequences of length w from the standardized traffic flow observation time series in the manner of a sliding window. Each window partially overlaps with the previous window to obtain several subsequences of a fixed length.

[0069] The standardized traffic flow observation time series is blocked according to a fixed window length w and a step size s, and subsequences centered on the time step t are extracted. The range of each subsequence is Among them, , and T is the total length of the time series data.

[0070] Through overlapping block processing, each subsequence contains some identical data points, so that the correlation between adjacent time windows is retained, and key time information will not be lost due to division.

[0071] Through the sliding window method, each time point may be utilized by multiple subsequences, increasing the number of effective training samples and enhancing the generalization ability of the model. Alleviate data sparsity: In the case of less data in some time periods, overlapping block processing can ensure that each time point will not be ignored, improving data coverage.

[0072] 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 dependence relationships.

[0073] Since multiple windows share some identical data points, the prediction results will be integrated during the calculation of multiple windows, thereby reducing the impact of sudden changes on the prediction.

[0074] Step 200: Construct a traffic flow time series prediction model.

[0075] Construct a traffic flow time series prediction model with reference 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 depthwise 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 depthwise separable convolution block, the output of the interactive depthwise 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.

[0076] 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 through layer normalization; the interactive depthwise 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 depthwise 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.

[0077] As a specific example, this step 200 may include the following sub-steps:

[0078] Step 210: Perform layer normalization and linear processing on the preprocessed traffic flow observation time series to obtain the linearly transformed traffic flow observation time series.

[0079] Layer normalization normalizes all features of each preprocessed traffic flow observation time series sample, rather than normalizing batch data. This helps to stabilize the network training process and is suitable for data with a large variation in batch size such as time series.

[0080] In this embodiment, a specific implementation manner of step 210 may be:

[0081] Step 2101: For the preprocessed traffic flow observation time series (with shape (B, T, D), where D is the feature dimension), layer normalization calculates the mean and standard deviation on the last dimension D:

[0082] ;

[0083] ;

[0084] 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.

[0085] Step 2102: Perform the first normalization on the preprocessed traffic flow observation time series to obtain the first-normalized traffic flow observation time series:

[0086] ;

[0087] where, represents the first-normalized traffic flow observation time; is a very small number used to prevent division-by-zero errors.

[0088] Step 2103: Perform a linear transformation on the first-normalized traffic flow observation time series:

[0089] ;

[0090] where, represents the traffic flow observation time after the linear transformation, represents the scaling parameter, represents the offset parameter.

[0091] The traffic flow observation time after the linear transformation Furthermore, the traffic flow observation time series after the linear transformation is obtained .

[0092] Step 220: Perform frequency-domain processing and feature extraction on the traffic flow observation time series after the linear transformation, and convert it back to the time domain to maintain the consistency of the data structure.

[0093] Use the fast Fourier transform to transform the traffic flow observation time series after the linear transformation into the frequency domain. Construct two learnable parameter matrices, and generate an adaptive weight matrix through the parameter matrices; the adaptive weight matrix is used to filter the frequency-domain signal, emphasizing the enhancement or suppression of certain frequency components to help the network pay more attention to important frequency components. Use the adaptive weight matrix to filter the frequency-domain signal, and the purpose of filtering is to emphasize the feature extraction of the high-frequency or low-frequency part.

[0094] At the same time, use the random masking strategy 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, fuse the frequency-domain features after filtering and random masking to obtain a more comprehensive feature representation.

[0095] In this embodiment, a specific implementation of step 220 may be as follows:

[0096] Step 2201: Use the fast Fourier transform to convert the linearly transformed traffic flow observation time series into the frequency domain to obtain the traffic flow observation time series after the fast Fourier transform:

[0097] ;

[0098] In the above formula, represents the traffic flow observation time series after linear transformation; represents the traffic flow observation time series after the fast Fourier transform, represents the number of frequency points; F is the frequency domain representation of the traffic flow observation time series after linear transformation.

[0099] Step 2202: Use the random masking strategy to adaptively adjust the traffic flow observation time series after the fast Fourier transform to enhance the robustness and diversity of the model.

[0100] Random frequency masking to generate a random binary masking matrix M, whose shape is the same as the frequency domain representation of the traffic flow observation time series after linear transformation, and the value of each element is 0 or 1, indicating whether to retain the frequency component.

[0101] ;

[0102] Among them, indicates that each frequency component has a 50% probability of being retained; indicates generating a binary number (0 or 1) by the Bernoulli distribution.

[0103] Multiply the traffic flow observation time series after the fast Fourier transform and the binary masking matrix M element by element according to the main elements to obtain the masked data:

[0104] ;

[0105] In the above formula, represents the masked data; ⊙ represents element-by-element multiplication according to the main elements.

[0106] Step 2203: Construct two learnable parameter matrices, and generate an adaptive weight matrix through the parameter matrices; 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. Filter the frequency domain signal using the adaptive weight matrix, and the purpose of filtering is to emphasize the feature extraction of the high-frequency or low-frequency part.

[0107] 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;

[0108] 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:

[0109] ;

[0110] Masking Data With learnable local complex weights Multiply them together to get the local complex masking feature:

[0111] ;

[0112] In the above formula, represents the complex weighted frequency domain traffic flow characteristics; Represents a local complex masking feature.

[0113] 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:

[0114] ;

[0115] In the above formula, Represents the complex fused frequency domain-masked traffic flow characteristics.

[0116] 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 :

[0117] ;

[0118] IFFT ensures that the enhanced features remain consistent with the original data structure of the input time series.

[0119] Step 230: performing layer normalization on the time domain signal after the frequency domain enhancement processing.

[0120] 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.

[0121] Step 240: Extract interactive features of the time-domain signal through an interactive depthwise separable convolution block and perform mapping to obtain a mapped output feature.

[0122] The convolution operation combines different receptive fields and can extract features from different time scales. After extracting features at multiple scales, a feature interaction mechanism is used to fuse features at different scales. By calculating the interaction between features at different scales, more complex temporal dependencies are captured.

[0123] In this embodiment, a specific implementation manner of step 240 may be:

[0124] Step 2401: The time-domain signal S is used as the input of the interactive depthwise separable convolution block, and the interactive depthwise separable convolution block includes a first group of depthwise separable convolution and a second group of depthwise separable convolution.

[0125] The first group of depthwise separable convolution:

[0126] The first group of depth convolution: ;

[0127] The first group of point convolution: ;

[0128] The first group of activation and Dropout: ;

[0129] The second group of depthwise separable convolution:

[0130] The second group of depth convolution: ;

[0131] The second group of point convolution: ;

[0132] The second group of activation and Dropout: ;

[0133] In the above formula, represents the ReLU activation function; represents depth convolution, represents point convolution; represents the output of the first group of depth convolution; represents the first group of point convolution; represents the output of the first group of activation and Dropout; represents the output of the second group of depth convolution; represents the second group of point convolution; represents the output of the second group of activation and Dropout.

[0134] Among them, the first group of depth convolutions and the second group of depth convolutions use depth convolutions with two different kernel sizes (3 and 5) to extract features in different receptive field ranges, and the point convolution uses one-dimensional convolution for cross-channel information mixing.

[0135] Step 2402: Feature interaction. Feature interaction is achieved through cross multiplication to obtain two interaction features.

[0136] The output of the first group of point convolutions is multiplied by the output of the second group of activation and Dropout to obtain the first interaction feature:

[0137] ;

[0138] Among them, represents the first interaction feature.

[0139] The second group of point convolutions is multiplied by the output of the first group of activation and Dropout to obtain the second interaction feature:

[0140] ;

[0141] Among them, represents the second interaction feature.

[0142] The first interaction feature and the second interaction feature are added together as the final interaction feature:

[0143] ;

[0144] Among them, represents the final interaction feature.

[0145] Step 2403: Map the final interaction feature back to the input feature dimension through ordinary convolution to obtain the mapped output feature:

[0146] ;

[0147] Among them, the convolution kernel size is 1×1; represents the mapped output feature.

[0148] Step 250: Further extract global temporal features through a multi-layer neural network and fuse the features extracted by each layer. This step aims to integrate information at different levels to obtain global temporal information;

[0149] The features output by the previous layer mapping are further extracted, global temporal features are extracted through a multi-layer neural network, and the features extracted by each layer are fused to integrate information at different levels, so as to obtain complete global temporal information.

[0150] Step 260: Perform a residual connection on the preprocessed traffic flow observation time series and the mapped output features output by the interactive depthwise separable convolution, and calculate the residual fusion features.

[0151] To ensure the integrity of the model during feature processing and prevent information loss, a residual connection mechanism is adopted. The preprocessed traffic flow observation time series is added to the mapped output features output by the interactive depthwise separable convolution network as follows:

[0152] ;

[0153] In the above formula, represents the residual fusion features.

[0154] Residual connections are crucial in the deep structure of the model. They can alleviate the degradation problem of information during multi-layer propagation, while retaining the key information of the original features and enhancing the learning stability of the model.

[0155] Step 270: Convert the residual fusion features into predicted values through a fully connected layer and perform denormalization to obtain the final prediction result.

[0156] In this embodiment, a specific implementation of Step 270 can be as follows:

[0157] Step 2701: Convert the residual fusion features into predicted values through a fully connected layer:

[0158] ;

[0159] Among them, represents the weight matrix, which determines the mapping from input features to output features; represents the bias vector, which is used to adjust the output value; represents the output of the linear layer.

[0160] Step 2702: Perform denormalization to obtain the final prediction result:

[0161] ;

[0162] In the above formula, represents the final prediction result. This result combines the features in the frequency domain and the time domain and can achieve excellent results in prediction accuracy and data pattern preservation.

[0163] Step 300: Train the traffic flow time series prediction model to obtain a trained traffic flow time series prediction model.

[0164] Step 400: Input the time series data to be predicted into the trained traffic flow time series prediction model to obtain a prediction result.

[0165] Refer to Figure 2 , where 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 when performing long-sequence traffic flow prediction, the present invention can also ensure accuracy and effective fitting effect.

[0166] 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 foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate 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 time series prediction method based on FFT and interactive depthwise separable convolution, characterized in that Including 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, including: Using the fast Fourier transform to transform the linearly transformed traffic flow observation time series to the frequency domain to obtain the fast Fourier transform of the traffic flow observation time series; randomly frequency masking to generate a random binary masking matrix; multiplying the fast Fourier transform of the traffic flow observation time series and the binary masking matrix element by element according to the main elements to obtain masked data; constructing learnable global complex weights and learnable local complex weights; weighting the frequency domain representation of the linearly transformed traffic flow observation time series by the learnable global complex weights to obtain complex weighted frequency domain traffic flow features: multiplying the masked data by the learnable local complex weights to obtain local complex masked features; integrating the complex weighted frequency domain traffic flow features and the local complex masked features to capture comprehensive frequency details to obtain complex fused frequency domain - masked traffic flow features: converting the complex fused frequency domain - masked traffic flow features back to the time domain by applying the inverse fast Fourier transform to obtain a time domain signal; Extracting the interaction features of the time domain signal through an interactive depth - separable convolution block and performing mapping to obtain the mapped output features; Performing residual connection on the preprocessed traffic flow observation time series and the mapped output features, calculating to obtain the residual fusion features; converting the residual fusion features to prediction values through a fully connected layer and performing denormalization to obtain the final prediction result.

2. The time series prediction method based on FFT and interactive depthwise separable convolution according to claim 1, characterized in that 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: Removing 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; Normalizing the historical traffic flow observation time series after mean removal to further obtain the preprocessed traffic flow observation time series.

3. A time series prediction method based on FFT and interactive depthwise separable convolution according to claim 1, characterized in that The performing layer normalization and linear processing on 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, calculating the mean and standard deviation; Based on the calculated mean and standard deviation, performing the first normalization on the preprocessed traffic flow observation time series to obtain the first - normalized traffic flow observation time series; Performing linear transformation on the first - normalized traffic flow observation time series to obtain the linearly transformed traffic flow observation time series.

4. A time series prediction method based on FFT and interactive depthwise separable convolution according to claim 3, characterized in that, The extracting the interaction features of the time domain signal through an interactive depth - separable convolution block and performing mapping to obtain the mapped output features includes: The time-domain signal is used as the input of the interactive depthwise separable convolution block, and the interactive depthwise separable convolution block includes depthwise separable convolution, pointwise convolution, activation, and Dropout; the interactive depthwise separable convolution block includes a first group of depthwise separable convolution and a second group of depthwise separable convolution; The output of the first group of pointwise convolutions is multiplied by the output of the second group of activation and Dropout to obtain a first interaction feature: the output of the second group of pointwise convolutions is multiplied by the output of the first group of activation and Dropout to obtain a second interaction feature: the first interaction feature and the second interaction feature are added together as the final interaction feature: The final interaction feature is mapped back to the input feature dimension through ordinary convolution to obtain a mapped output feature.

5. A time series prediction method based on FFT and interactive depthwise separable convolution according to claim 4, characterized in that The process of converting the residual fusion feature into a predicted value through a fully connected layer and performing denormalization to obtain the final prediction result includes: The residual fusion feature is converted into a predicted value through a fully connected layer: ； Among them, represents the weight matrix; represents the bias vector, which is used to adjust the output value; represents the output of the linear layer; represents the residual fusion feature; The output of the linear layer is denormalized to obtain the final prediction result: ； Among them, represents the final prediction result; represents the mean of the time series; represents the standard deviation of the time series.

6. The time series prediction method based on FFT and interactive depthwise separable convolution according to claim 1, characterized in that 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.

Images