A Parking Space Status Prediction Method Based on Frequency Awareness and Hybrid Experts

Abstract

This invention provides a parking space status prediction method based on frequency perception and hybrid expert analysis, belonging to the field of data processing technology. Specifically, it includes: extracting time-domain features; converting the time-domain features to the frequency domain, extracting the highest frequency point, performing frequency domain masking and recombination to obtain the highest frequency component, then repeating the frequency domain masking and recombination M-1 times for the frequency components of the mask to obtain M frequency components; calculating the similarity between the remaining M-1 frequency components and the highest frequency component, and calculating the weight of each frequency component accordingly; inputting the highest frequency component into the corresponding expert module in the time domain after undergoing inverse discrete cosine transform; inputting the remaining frequency components into the corresponding expert modules in sequence according to their corresponding weights after undergoing inverse discrete cosine transform; and fusing the M prediction results according to the weights of their corresponding frequency components to obtain the final parking space status prediction result. This invention improves prediction efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a parking space status prediction method based on frequency perception and hybrid experts. Background Technology

[0002] Currently, with the improvement of residents' consumption levels and the rapid growth in the number of cars, parking has become a major challenge in urban management. Traditional parking lot management models lead to unreasonable allocation of parking resources and low efficiency. Existing prediction methods struggle to effectively identify and utilize multi-frequency, especially high-frequency, features in parking space status data. Furthermore, these methods typically employ a uniform modeling approach to handle features of different frequencies, resulting in insufficient model perception of multi-frequency features, thus affecting prediction performance.

[0003] It is evident that there is an urgent need for a parking space status prediction method based on frequency perception and hybrid experts with high prediction efficiency and accuracy. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a parking space status prediction method based on frequency perception and hybrid experts, which at least partially solves the problems of poor prediction efficiency and accuracy in the prior art.

[0005] This invention provides a parking space status prediction method based on frequency sensing and hybrid experts, including:

[0006] Step 1: Collect historical parking space data and normalize it, then extract time-domain features based on this data;

[0007] Step 2: Use discrete cosine transform to convert the time domain features to the frequency domain, extract the highest frequency point, and then perform frequency domain masking and recombination to obtain the highest frequency component. Then repeat frequency domain masking and recombination M-1 times for the frequency components of the mask to obtain M frequency components. Calculate the similarity between the last M-1 frequency components and the highest frequency component, and calculate the weight of each frequency component accordingly.

[0008] Step 3: Input the highest frequency component into the corresponding expert module in the time domain after inverse discrete cosine transform, and input the remaining frequency components into the corresponding expert modules in order according to their corresponding weights after inverse discrete cosine transform, so as to obtain the prediction results of M expert modules.

[0009] Step 4: Fuse the M prediction results according to the weights of their corresponding frequency components to obtain the final parking space status prediction result.

[0010] According to a specific implementation of an embodiment of the present invention, step 2 specifically includes:

[0011] Step 2.1: Use Discrete Cosine Transform to convert the time-domain features into frequency-domain features.

[0012] ;

[0013] in, For frequency domain characteristics, It is a time-domain feature. is the sequence length of the time-domain feature, n is the sampling point index of the time-domain sequence, and k is the frequency component index of the frequency-domain feature;

[0014] Step 2.2: Use a high-pass filter to extract the highest frequency point in the frequency domain features, then use a mask encoder to mask it and the frequency components other than the low-frequency components, and reassemble the remaining unmasked frequency components to obtain a new frequency component. The low-frequency component is a set of frequency points in the frequency domain that are close to the zero frequency within a preset range.

[0015] Step 2.3: Repeat step 2.2 for the frequency components of the mask to obtain the new frequency components and new frequency elements of the mask;

[0016] Step 2.4: Repeat steps 2.2 and 2.3 M-2 times to obtain M frequency components;

[0017] Step 2.5: Calculate the Euclidean distance between the last M-1 frequency components and the first highest frequency component as the similarity. Substitute the similarity into the preset formula to obtain the weight of each frequency component.

[0018] According to a specific implementation of an embodiment of the present invention, the expression of the preset formula is as follows:

[0019] ;

[0020] in, This represents the weight of the i-th frequency component. σ is the Euclidean distance between the i-th frequency component and the highest frequency component j, and σ is a hyperparameter used to control the influence of the distance on the weights.

[0021] According to a specific implementation of the present invention, the expression for the inverse discrete cosine transform is:

[0022] ;

[0023] in, It is the time-domain value of the i-th frequency component output at the n-th sampling point. It is the value of the k-th point on the i-th frequency component.

[0024] According to one specific implementation of the present invention, the expert module includes a deep convolutional module and a multi-order KAN network.

[0025] According to a specific implementation of an embodiment of the present invention, step 3 specifically includes:

[0026] Step 3.1: For the time-domain data corresponding to the highest frequency component and the time-domain data corresponding to each remaining frequency component, convolution is performed using convolution kernels of different sizes in the depthwise convolution module. Additionally, for the time-domain data corresponding to the highest frequency component and the time-domain data corresponding to each remaining frequency component input in weight order, time-domain data is captured using Chebyshev polynomials of different orders in a multi-order KAN network. Characteristics of different frequencies;

[0027] Step 3.2: Combine the outputs of each convolution kernel and each order of Chebyshev polynomial to obtain M prediction results.

[0028] According to a specific implementation of an embodiment of the present invention, the expression for the parking space status prediction result is as follows:

[0029] ;

[0030] in, It is the first The weights of each frequency component, It is the prediction result of the i-th expert module.

[0031] The parking space status prediction scheme based on frequency perception and hybrid experts in this embodiment of the invention includes: Step 1, collecting historical parking space data and normalizing it, and extracting time-domain features accordingly; Step 2, using discrete cosine transform to convert the time-domain features to the frequency domain, extracting the highest frequency point, performing frequency domain masking and recombination to obtain the highest frequency component, and then repeating frequency domain masking and recombination M-1 times for the frequency components of the mask to obtain M frequency components, calculating the similarity between the last M-1 frequency components and the highest frequency component, and calculating the weight of each frequency component accordingly; Step 3, inputting the highest frequency component after inverse discrete cosine transform into the corresponding expert module in the time domain, and inputting the remaining frequency components after inverse discrete cosine transform into the corresponding expert modules in sequence according to their corresponding weights to obtain the prediction results of M expert modules; Step 4, fusing the M prediction results according to the weights of their corresponding frequency components to obtain the final parking space status prediction result.

[0032] The beneficial effects of this invention are as follows: By introducing frequency perception and similarity analysis methods to address different frequency characteristics in parking space status data, the present invention first transforms the time-domain features to the frequency domain using discrete cosine transform, and then extracts the highest frequency point using a high-pass filter. Next, the frequency components are masked to obtain M frequency components with different frequency representations. Then, the similarity between each frequency component and the highest frequency component is calculated using Euclidean distance, and the weights of each frequency component are adaptively adjusted based on this similarity. Then, the time-series data of the high-frequency and remaining frequency components are input into specially designed hybrid expert modules for processing. Each expert module consists of a deep convolutional network and a multi-order KAN network, capable of extracting the most relevant features from the time-series data of different frequencies. Finally, the prediction results of each expert module are weighted and fused according to the weights of each frequency component to obtain the final parking space status prediction, improving prediction efficiency and accuracy. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart illustrating a parking space status prediction method based on frequency sensing and hybrid experts provided in an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram illustrating the principle of a frequency sensing and similarity analysis process provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the structure of a hybrid expert module combining multi-level KAN and deep convolution provided in an embodiment of the present invention. Detailed Implementation

[0037] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0038] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0040] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0041] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0042] In recent years, with the acceleration of urbanization and the continuous increase in the number of motor vehicles, the problem of parking difficulties in cities has become increasingly prominent, becoming one of the major bottlenecks restricting the development of modern urban transportation. Against this backdrop, intelligent transportation systems (ITS) have emerged, aiming to optimize the allocation and utilization of transportation resources and improve traffic efficiency through advanced information technology. Parking space status prediction, as a key component of ITS, is of paramount importance in alleviating parking difficulties. The core of parking space status prediction lies in analyzing historical parking space occupancy data to uncover its spatiotemporal distribution patterns, thereby achieving accurate predictions of future parking space status. This not only provides real-time parking guidance for drivers but also offers decision support for parking lot managers and urban transportation planners. However, most existing parking space status prediction methods have limitations and struggle to meet the high accuracy and robustness requirements of practical applications.

[0043] Traditional parking space status prediction methods primarily rely on time series analysis and machine learning techniques, such as linear regression and support vector machines. These methods often treat time-series data as a whole, neglecting the information contained in different frequency components. In reality, parking space status data possesses complex spatiotemporal characteristics, including fluctuations of varying frequencies, such as real-time weather, temporary fluctuations caused by accidents, medium-term diurnal variation, and long-term seasonal trends. These different frequency components reflect different patterns and influencing factors in parking space status changes, and their comprehensive consideration and effective utilization are crucial for improving prediction accuracy. Furthermore, existing methods lack targeted modeling strategies when dealing with high-frequency and other frequency components. High-frequency components typically capture rapid changes in parking space status, reflecting short-term dynamic demand fluctuations; while low-frequency components reflect long-term trends and periodic patterns. Applying a uniform modeling approach to all frequency components makes it difficult to fully leverage their individual characteristics and advantages, thus limiting the performance of the prediction model.

[0044] This invention provides a parking space status prediction method based on frequency perception and hybrid experts, which can be applied to the parking space status monitoring process in urban management scenarios.

[0045] See Figure 1 This is a flowchart illustrating a parking space status prediction method based on frequency sensing and hybrid experts, provided by an embodiment of the present invention. Figure 1 As shown, the method mainly includes the following steps:

[0046] Step 1: Collect historical parking space data and normalize it, then extract time-domain features based on this data;

[0047] In practice, the time-series data of parking space status is first normalized, normalizing the parking space status data (such as occupancy rate, vacancy rate, etc.) to the range [0, 1]. After processing, we obtain the normalized time-series data, which is denoted as... This includes state data for each time point from time 1 to n, i.e.: ={ , ,..., }

[0048] Then, feature extraction is performed on the normalized data to extract the time-domain features of the data (hours, weeks, days, real-time weather, traffic accidents, etc.) in order to capture the periodic changes in parking space status and the features at different time scales, ensuring the consistency and usability of the input data.

[0049] Step 2: Use discrete cosine transform to convert the time domain features to the frequency domain, extract the highest frequency point, and then perform frequency domain masking and recombination to obtain the highest frequency component. Then repeat frequency domain masking and recombination M-1 times for the frequency components of the mask to obtain M frequency components. Calculate the similarity between the last M-1 frequency components and the highest frequency component, and calculate the weight of each frequency component accordingly.

[0050] In practice, the discrete cosine transform (DCT) can be used to transform the time-domain features to the frequency domain.

[0051] ;

[0052] in, Frequency characteristics It is a time-domain feature. is the sequence length of the time-domain feature, n is the sampling point index of the time-domain sequence, and k is the frequency component index of the frequency-domain feature.

[0053] Then, after the high-pass filter extracts the highest frequency point, the original frequency is passed through a low-pass filter to obtain the remaining frequency components. Simultaneously, the highest frequency point is extracted, and the remaining frequency components are masked. Specifically, all frequency components except the highest frequency point and low-frequency components are masked. Here, low-frequency components refer to the set of frequency points close to zero frequency in the frequency domain. Next, the retained high-frequency and low-frequency components are recombine to form a new frequency component. To further obtain frequency characteristics at different levels, the above masking and recombination operation is repeated M-1 times for the masked component, each time selecting a different highest frequency point and low-frequency portion for recombination, ultimately obtaining M frequency components with different frequency characteristics.

[0054] Then, the remaining frequency components There are M-1 such sets, each capable of focusing on high-frequency identification of various frequency components. The set of all frequency components can be expressed as a formula:

[0055] ={ ,…, ,…, };

[0056] Where M-1 represents the number of frequency components. ∈ Let F represent the k-th point of the i-th frequency component, F represent the dimension of the frequency component, i.e. the number of features of each frequency component, and n represent the time step of each signal segment, i.e. the number of samples contained in the data segment.

[0057] For the remaining M-1 frequency components, a method such as Euclidean distance can be used to measure the similarity between frequency component i and high-frequency component j. The specific formula is as follows:

[0058] ;

[0059] here, and Let represent the i-th frequency component and the high-frequency component j, respectively. L is the number of dimensions of the frequency components (i.e., the number of frequency points for each frequency component). k is the index used to traverse each frequency component.

[0060] If frequency signal If a signal is very similar in amplitude to a high-frequency signal, then the Euclidean distance between them will be small, indicating that... At this frequency point, it is similar to a high-frequency signal. If their amplitudes differ significantly, then the Euclidean distance will be large, indicating... The difference between high-frequency signals and high-frequency signals is significant at this frequency point. In this way, comparisons can be made frequency-by-frequency. The amplitude differences between high-frequency components are compared, and their similarity is measured using Euclidean distance.

[0061] Next, based on the calculated Euclidean distance, their weights can be adaptively calculated. Specifically, frequency components with smaller Euclidean distances to higher-frequency components are considered to be of the same frequency and receive higher weights, while those with larger distances receive lower weights. A new weight is assigned to each frequency component using the following formula:

[0062] ;

[0063] in, It is the weight of the i-th frequency component; σ is the Euclidean distance between the i-th frequency component and the high-frequency component j; σ is a hyperparameter that controls the influence of the distance on the weights.

[0064] The weight of frequency component i is related to its distance from high-frequency component j; components that are closer will have a greater influence on the weight of i. After obtaining M-1 weights, they are sorted, with larger weights indicating that it is a high-frequency component.

[0065] After calculating the weights in the frequency domain, the inverse discrete cosine transform (IDCT) is used to transform the frequency components. Convert back to time domain This is so that it can be further input into the MoE module:

[0066] ;

[0067] in, It is the time domain value of the i-th frequency component output at the n-th sampling point; It is the value of the k-th point on the i-th frequency component.

[0068] Step 3: Input the highest frequency component into the corresponding expert module in the time domain after inverse discrete cosine transform, and input the remaining frequency components into the corresponding expert modules in order according to their corresponding weights after inverse discrete cosine transform, so as to obtain the prediction results of M expert modules.

[0069] In practice, the highest frequency component can be... After being transformed to the time domain by inverse discrete cosine transform, the data is input into the corresponding high-frequency processing expert module. Furthermore, based on the obtained weights for each segment, each segment of time-series data is processed. The data is input into its respective expert module. In short, all time-series data is fed into its corresponding expert module to achieve personalized processing for high-frequency and other different components. Based on this, each expert module independently processes its corresponding time-series data. The expert module consists of a parallel architecture composed of deep convolutional layers and multi-level KAN networks, designed to extract the most relevant features from high-frequency and other different frequency time-series data.

[0070] The deep convolution module uses multiple convolutional layers to process time-series data. Each convolutional layer adapts to time-series data of different frequencies using convolutional kernels of varying sizes, thereby performing specific modeling and learning for different frequency components. For example, for high-frequency components, smaller convolutional kernels can be used to focus on local, short-term data changes to capture short-term fluctuations, thus better adapting to the characteristics of different frequency components. For low-frequency components, larger convolutional kernels can be used to cover a wider data range and a longer time span, thereby smoothing out short-term fluctuations and noise to capture long-term trends.

[0071] M-order KAN networks utilize multiple Chebyshev polynomials of different orders to capture input data. When fitting data to different frequencies, low-order Chebyshev polynomials, being simple functions, automatically ignore high-frequency fluctuations and focus primarily on the overall trend. This allows them to effectively smooth out short-term noise and fluctuations, extracting long-term, smooth trend information when processing non-high-frequency components. For example, a first-order Chebyshev polynomial is a linear function, exhibiting good fitting ability for long-term, slowly changing trends. As the order of the Chebyshev polynomial increases, its functional form becomes more complex, leading to larger variations over a smaller time frame. This makes them well-suited for capturing short-term, rapidly changing high-frequency components in signals.

[0072] Step 4: Fuse the M prediction results according to the weights of their corresponding frequency components to obtain the final parking space status prediction result.

[0073] In practice, after obtaining the prediction results from all expert modules, these prediction results are weighted and fused with the previously calculated weights of the corresponding frequency components to obtain the final prediction result. The purpose of weighted fusion is to obtain an accurate, comprehensive prediction result that highlights high-frequency components. Specifically, the higher the weight of a frequency component, the greater the proportion of its corresponding expert module's prediction result in the final prediction. The specific formula is as follows:

[0074] ;

[0075] in, It is the weight of the i-th frequency component. This is the prediction result of the i-th expert module. In this way, we can fuse the information from different frequency components to produce a prediction that combines the features of all frequency components.

[0076] By employing the above method, this invention fully utilizes the time-series data characteristics of different frequency components. Through a combination of frequency domain sensing, frequency similarity analysis, and a hybrid expert module, it achieves high-precision parking space status prediction. The prediction results output by the hybrid expert module reflect the comprehensive parking space status. This has a positive promoting effect on optimizing parking space resource allocation in complex parking environments and improving the parking experience for drivers.

[0077] This embodiment provides a parking space status prediction method based on frequency perception and hybrid expert analysis. By introducing frequency perception and similarity analysis methods to address different frequency features in the parking space status data, the method first transforms the time-domain features to the frequency domain using discrete cosine transform (DCT), and then extracts the highest frequency point using a high-pass filter. Next, the frequency components are masked to obtain M frequency components with different frequency representations. Then, the similarity between each frequency component and the highest frequency component is calculated using Euclidean distance, and the weights of each frequency component are adaptively adjusted based on this similarity. Then, the time-series data of the high-frequency and remaining frequency components are input into specially designed hybrid expert modules for processing. Each expert module consists of a deep convolutional network and a multi-order KAN network, capable of extracting the most relevant features from the time-series data of different frequencies. Finally, the prediction results of each expert module are weighted and fused according to the weights of each frequency component to obtain the final parking space status prediction, improving prediction efficiency and accuracy.

[0078] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0079] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A parking space status prediction method based on frequency sensing and hybrid experts, characterized in that, include: Step 1: Collect historical parking space data and normalize it, then extract time-domain features based on this data; Step 2: Use discrete cosine transform to convert the time domain features to the frequency domain, extract the highest frequency point, and then perform frequency domain masking and recombination to obtain the highest frequency component. Then repeat frequency domain masking and recombination M-1 times for the frequency components of the mask to obtain M frequency components. Calculate the similarity between the last M-1 frequency components and the highest frequency component, and calculate the weight of each frequency component accordingly. Step 2 specifically includes: Step 2.1: Use Discrete Cosine Transform to convert the time-domain features into frequency-domain features. ； in, For frequency domain characteristics, It is a time-domain feature. is the sequence length of the time-domain feature, n is the sampling point index of the time-domain sequence, and k is the frequency component index of the frequency-domain feature; Step 2.2: Use a high-pass filter to extract the highest frequency point in the frequency domain features, and then use a mask encoder to mask the frequency components other than the highest frequency point and low frequency components. Then, reassemble the retained high frequency components and low frequency components to obtain a new frequency component. The low frequency component is a set of frequency points in the frequency domain that are close to the zero frequency within a preset range. Step 2.3: Repeat step 2.2 for the frequency components of the mask to obtain the new frequency components and new frequency elements of the mask; Step 2.4: Repeat steps 2.2 and 2.3 M-2 times to obtain M frequency components; Step 2.5: Calculate the Euclidean distance between the last M-1 frequency components and the first highest frequency component as the similarity. Substitute the similarity into a preset formula to obtain the weight of each frequency component. The expression of the preset formula is: ； in, This represents the weight of the i-th frequency component. It is the Euclidean distance between the i-th frequency component and the first highest frequency component j, and σ is a hyperparameter used to control the influence of the distance on the weight; Step 3: The highest frequency component is input into the corresponding expert module in the time domain after undergoing inverse discrete cosine transform. The remaining frequency components are input into the corresponding expert modules in order according to their weights after undergoing inverse discrete cosine transform, so as to obtain the prediction results of M expert modules. The expert modules include a deep convolution module and a multi-order KAN network. Step 3 specifically includes: Step 3.1: For the time-domain data corresponding to the highest frequency component and the time-domain data corresponding to each remaining frequency component, convolution is performed using convolution kernels of different sizes in the depthwise convolution module. Additionally, for the time-domain data corresponding to the highest frequency component and the time-domain data corresponding to each remaining frequency component input in weight order, time-domain data is captured using Chebyshev polynomials of different orders in a multi-order KAN network. Characteristics of different frequencies; Step 3.2: Combine the outputs of each convolution kernel and the Chebyshev polynomial of each order to obtain M prediction results; Step 4: Fuse the M prediction results according to the weights of their corresponding frequency components to obtain the final parking space status prediction result.

2. The method according to claim 1, characterized in that, The expression for the inverse discrete cosine transform is: ； in, It is the time-domain value of the i-th frequency component output at the n-th sampling point. It is the value of the k-th point on the i-th frequency component.

3. The method according to claim 2, characterized in that, The expression for the parking space status prediction result is as follows: ； in, It is the first The weights of each frequency component, It is the prediction result of the i-th expert module.

Images