Trout demand prediction system and method based on multi-scale data fusion

Abstract

The invention provides a text travel demand prediction system and method based on multi-scale data fusion, and relates to the technical field of data processing, and the system comprises a module for obtaining text travel historical time sequence data; the module is used for extracting time-frequency features of the historical time sequence data of the text travel, splicing the time-frequency features to obtain spliced features and fusing the spliced features to obtain fused features; the module is used for stacking the multi-scale division results of the fusion features to obtain multi-scale embedding; a module for performing inter-variable and intra-variable feature interaction on multi-scale embedding to obtain target multi-scale embedding; the module is used for carrying out weighted fusion on the obtained splicing features, fusion features and target multi-scale embedding to obtain fusion embedding; the module is used for performing residual connection on fusion embedding and text travel historical time sequence data to obtain target embedding; and the output projection layer is used for processing the target embedding to obtain a text travel demand prediction result. The objective of the invention is to improve travel demand prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a system and method for predicting cultural and tourism demand based on multi-scale data fusion. Background Technology

[0002] Time series forecasting is one of the core technologies in the era of big data. In the cultural tourism industry, with the construction of various digital service platforms and other projects, there is a need for high-precision, multi-scale accurate forecasting of the massive amounts of diverse operational indicators (such as daily visitor traffic, sales of cultural and creative products, and system user concurrency) generated by these platforms. This is crucial for achieving efficient and scientific management of visitor profiling, targeted promotion, and the integration of cultural tourism consumption with e-commerce subsystems.

[0003] However, cultural and tourism operation data differs from traditional data, exhibiting the following characteristics that pose significant challenges to existing forecasting models. These characteristics include: coexistence of multi-scale time series patterns (the series simultaneously contains intraday fluctuations, weekly / monthly / quarterly periodicity, and annual macroeconomic trends); high coupling of multiple variables (different indicators on the platform (such as ticket revenue and accommodation bookings) influence each other, resulting in complex coupling relationships); and complex noise and periodicity (affected by external factors (weather, holidays, and emergencies), the periodic patterns in the series are complex and contain high-frequency noise). Currently used time series forecasting methods cannot accurately predict long-term series data of cultural and tourism operation data that simultaneously possesses the above characteristics. Summary of the Invention

[0004] In view of this, this application provides a system and method for predicting cultural and tourism demand based on multi-scale data fusion. It aims to solve or partially solve the problems existing in the background technology.

[0005] The first aspect of this application provides a cultural tourism demand forecasting system based on multi-scale data fusion, the system comprising: The module includes a data acquisition module, a multi-resolution processing module, a multi-scale time series embedding module, a multi-stage feature fusion module, a feature fusion enhancement module, a residual connection module, and an output projection layer. The data acquisition module is used to acquire the cultural and tourism historical time series data of the target area. The cultural and tourism historical time series data is a three-dimensional tensor structure, which includes batch size, review window length and number of variables. The multi-resolution processing module is used to extract multi-resolution time-frequency features from the cultural and tourism historical time series data in parallel in the time domain and frequency domain, and to stitch them together to obtain stitched features, and to fuse the stitched features in the time dimension to obtain corresponding fused features. The multi-scale time series embedding module is used to perform multi-scale patching on the fused features and stack the patching results to obtain multi-scale embedding. The multi-stage feature fusion module is used to perform inter-variable feature interaction and intra-variable feature exchange processing on the multi-scale embedding to obtain the corresponding target multi-scale embedding. The feature fusion enhancement module is used to perform weighted fusion of the obtained spliced ​​features, the fused features and the target multi-scale embedding to obtain the corresponding fused embedding. The residual connection module is used to perform residual connection on the fused embedding and the cultural tourism historical time series data to obtain the target embedding; The output projection layer is used to process the target embedding and predict the corresponding cultural and tourism demand prediction results.

[0006] The second aspect of this application provides a method for predicting cultural and tourism demand based on multi-scale data fusion, applied to the cultural and tourism demand prediction system based on multi-scale data fusion described in the first aspect of this application. The method includes: The data acquisition module acquires the cultural and tourism historical time series data of the target area. The cultural and tourism historical time series data is a three-dimensional tensor structure, which includes batch size, review window length, and number of variables. The multi-resolution processing module extracts multi-resolution time-frequency features from the historical time series data of cultural tourism in parallel in the time and frequency domains, and splices them to obtain spliced ​​features. Furthermore, the spliced ​​features are fused in the time dimension to obtain corresponding fused features. The fused features are divided into multi-scale patches using a multi-scale time series embedding module, and the partitioning results are stacked to obtain multi-scale embeddings. The multi-stage feature fusion module performs inter-variable feature interaction and intra-variable feature exchange processing on the multi-scale embedding to obtain the corresponding target multi-scale embedding. The obtained spliced ​​features, fused features and target multi-scale embedding are weighted and fused by the feature fusion enhancement module to obtain the corresponding fused embedding. The target embedding is obtained by performing a residual connection between the fused embedding and the original input features through a residual connection module. The target embedding is processed by the output projection layer to predict the corresponding cultural and tourism demand.

[0007] The cultural tourism demand forecasting system based on multi-scale data fusion provided in this application has the following advantages: This application provides a cultural tourism demand prediction system based on multi-scale data fusion. The system comprises: a data acquisition module for acquiring historical time-series data of cultural tourism in a target area. The historical time-series data is a three-dimensional tensor structure, including batch size, review window length, and number of variables; a multi-resolution processing module for extracting multi-resolution time-frequency features from the historical time-series data in parallel in the time and frequency domains, stitching them together to obtain stitched features, and fusing the stitched features along the time dimension to obtain corresponding fused features; and a multi-scale time-series embedding module for processing the fused features. The system employs a multi-scale patch partitioning method, stacking the partitioning results to obtain multi-scale embeddings. A multi-stage feature fusion module performs inter-variable feature interaction and intra-variable feature exchange processing on the multi-scale embeddings to obtain the corresponding target multi-scale embedding. A feature fusion enhancement module performs weighted fusion of the obtained spliced ​​features, fused features, and target multi-scale embeddings to obtain the corresponding fused embedding. A residual connection module performs residual connection between the fused embedding and the historical time-series data of cultural tourism to obtain the target embedding. An output projection layer processes the target embedding and predicts the corresponding cultural tourism demand prediction results.

[0008] The core of this application lies in providing a cultural tourism demand forecasting system for digital cultural tourism platforms, enabling multi-scale prediction of diverse operational indicator sequence data. The system's core is an end-to-end prediction model based on a Patch Fusion MLP (Multilayer Perceptron) framework, designed to overcome current challenges in processing cultural tourism operational data, such as multi-scale patterns, multivariate variable coupling, and complex time-frequency feature fusion. Its processing flow consists of three main modules: a multi-resolution processing module, a multi-scale time series embedding module, and a multi-stage feature fusion module. For the multi-resolution processing module, this application, for the first time, systematically extracts and effectively fuses multi-resolution features simultaneously in the time domain (trend, seasonality) and frequency domain (periodicity, noise) within an MLP-based framework, addressing the weakness of existing methods in handling complex periodicity and noise. For the multi-scale time series embedding module, by processing patches of different lengths in parallel, it breaks the limitation of single-patch scale in current long-term data prediction, achieving simultaneous capture of multi-granularity time-series patterns of macro trends and micro details in cultural tourism operational data. The multi-stage feature fusion module not only performs the interaction between time steps of the sequence (traditional time-domain modeling), but more importantly, it explicitly models the complex coupling relationship between different operational indicator sequences, which improves the accuracy of the prediction model when processing multivariate sequences such as cultural and tourism consumption data, thereby effectively improving the accuracy of prediction. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram illustrating a cultural tourism demand prediction system based on multi-scale data fusion, as shown in one embodiment of this application. Figure 2 This is a structural diagram illustrating a cultural tourism demand prediction system based on multi-scale data fusion, as shown in one embodiment of this application. Figure 3 This is a structural diagram of a feature interaction module between variables in a cultural tourism demand prediction system based on multi-scale data fusion, as shown in one embodiment of this application. Figure 4 This is a structural diagram of an intravariable feature exchange module in a cultural tourism demand prediction system based on multi-scale data fusion, as shown in one embodiment of this application. Figure 5 This is a flowchart illustrating a method for predicting cultural and tourism demand based on multi-scale data fusion, as shown in one embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] refer to Figure 1 , Figure 1 This is a schematic diagram illustrating a cultural tourism demand prediction system based on multi-scale data fusion, as shown in one embodiment of this application. Figure 1As shown, the system includes: a data acquisition module 101, a multi-resolution processing module 102, a multi-scale time series embedding module 103, a multi-stage feature fusion module 104, a feature fusion enhancement module 105, a residual connection module 106, and an output projection layer 107. The data acquisition module is used to acquire historical time series data of cultural tourism in the target area. The historical time series data is a three-dimensional tensor structure, including batch size, review window length, and number of variables. The multi-resolution processing module is used to extract multi-resolution time-frequency features from the historical time series data of cultural tourism in parallel in the time and frequency domains, and to stitch them together to obtain stitched features. Furthermore, it performs time-dimensional fusion on the stitched features to obtain corresponding fused features. The multi-scale time series embedding module is used to perform multi-scale patching on the fused features and stack the patching results to obtain multi-scale embeddings; the multi-stage feature fusion module is used to perform inter-variable feature interaction and intra-variable feature exchange processing on the multi-scale embeddings to obtain the corresponding target multi-scale embeddings; the feature fusion enhancement module is used to perform weighted fusion on the obtained spliced ​​features, fused features and target multi-scale embeddings to obtain the corresponding fused embeddings; the residual connection module is used to perform residual connection on the fused embeddings and the cultural tourism historical time series data to obtain the target embeddings; the output projection layer is used to process the target embeddings and predict the corresponding cultural tourism demand prediction results.

[0013] In this application, the cultural and tourism historical time series data of the target area includes at least: multi-variable historical time series data of the target area containing multi-dimensional indicators of visitor flow and various consumption over a past period; the obtained cultural and tourism demand forecast results are a series of predicted values ​​of visitor flow and various consumption over a future period.

[0014] In this embodiment, since cultural tourism demand is influenced by geographical location and the characteristics of local tourist attractions, the cultural tourism demand prediction system based on multi-scale data fusion provided in this application first divides the prediction area when predicting cultural tourism demand. This division is based on a predetermined area range, and then, within that area, tourist attractions with similar characteristics are grouped into a single analysis cluster for cultural tourism demand prediction. For example, if a region has a large number of tourist attractions, it is first divided by city (or province, or a smaller area than a city) to obtain independent analysis areas at the city level. Then, tourist attractions with similar characteristics within these analysis areas are grouped into a single analysis cluster. If this analysis area includes attractions a, b, and c, and all three are natural landscapes, then the historical time-series data of cultural tourism within a 5-kilometer radius of each of these three attractions are acquired and used as a unified dataset for cultural tourism demand prediction.

[0015] In this embodiment, tourist attractions with similar characteristics within a single analysis area are grouped into a general analysis cluster and defined as a target area. The historical time-series data of the cultural and tourism sector within this target area is acquired through a data acquisition module. This historical time-series data has a three-dimensional tensor structure, including batch size, review window length, and the number of variables. The number of variables represents the number of cultural and tourism data types included in the historical time-series data. Specifically, the historical time-series data of the cultural and tourism sector within the target area includes at least: multivariate historical time-series data of multi-dimensional indicators of visitor flow and various consumption patterns within the target area over a past period. In multivariate time-series prediction, the objective of this application is to predict the future performance of a given historical time series. Below, predict time series , t represents the time step, B represents the batch size, L represents the lookback window length, T represents the prediction length, and C represents the number of variables.

[0016] In this embodiment, after the data acquisition module obtains the historical time-series data of the target region's cultural and tourism sector, the multi-resolution processing module performs multi-resolution time-frequency feature extraction on the acquired historical time-series data in both the time and frequency domains in parallel. Thus, multi-resolution time-domain and frequency-domain features can be obtained from the acquired historical time-series data. These two types of features are directly concatenated to obtain the corresponding concatenated features. Then, the concatenated features are further enhanced and fused using the time dimension to obtain the corresponding fused features. For example... Figure 2 As shown, Figure 2 The MRP Block in the code is a multi-resolution processing module.

[0017] In this embodiment, after obtaining the enhanced fusion feature through the multi-resolution processing module, the fusion feature is divided into multiple scale patches by the multi-scale time series embedding module to obtain multiple partitioning results at different patch scales. These partitioning results are then stacked to obtain a multi-scale embedding corresponding to the fusion feature. This multi-scale embedding includes embedding representations at multiple different patch scales corresponding to the fusion feature. For example... Figure 2 As shown, Figure 2 MTSEM in the text is a multi-scale time series embedding module.

[0018] In this embodiment, after obtaining the multi-scale embedding corresponding to the fused features through the processing of the multi-scale time series embedding module, the multi-stage feature fusion module performs feature interaction processing between different variables on the multi-scale embedding, and performs feature exchange processing within the same variable on the obtained processing results to obtain the corresponding target multi-scale embedding.

[0019] In this embodiment, after the aforementioned MLP-based multi-stage feature fusion, the prediction model obtains rich feature representations from the time-frequency domain, multi-scale embeddings, and inter-variable interactions. To effectively integrate these heterogeneous features from different sources and processing stages, and to ensure stable information flow, this application designs a feature fusion enhancement module in the prediction model. This module, through a composite feature fusion strategy and subsequent residual connections, achieves adaptive integration and information enhancement of multi-path features, providing the model with rich and stable feature representations. Since the model in this application contains multiple parallel feature extraction paths, each path captures different characteristics and patterns of the time series. The goal of this feature fusion enhancement module is to fully utilize this feature diversity, achieving information complementarity through a weighted fusion strategy. This ensures that the model can utilize both the periodic information in the time-frequency domain and capture multi-scale temporal dependencies, while maintaining sensitivity to the original temporal structure. This fusion strategy effectively avoids the limitations of a single feature source, improves the model's ability to model complex temporal patterns, and thus effectively improves the accuracy of subsequent prediction results. The expression for the multi-path weighted fusion strategy used in this feature fusion enhancement module is: ,in, For integration and embedding, To integrate features, For splicing features, Multi-scale embedding for the target This is the scaling factor for the multi-scale time series embedding module, used to control the contribution weight of multi-scale embedding features. Adjusting this parameter ensures the effective utilization of multi-scale features while avoiding the risk of overfitting due to excessive reliance. Figure 2 As shown, Figure 2 The MLP Block in the text is a multi-stage feature fusion module.

[0020] After obtaining the target multi-scale embedding through the multi-stage feature fusion module, the feature fusion enhancement module performs weighted fusion of the spliced ​​features, fused features, and target multi-scale embedding obtained from the above modules to obtain the corresponding fused embedding. For example... Figure 2 As shown, Figure 2 The FE Block in the text is the feature fusion enhancement module.

[0021] In this embodiment, the original cultural and tourism historical time series data may lose its original time series information after being processed by the multi-resolution processing module, the multi-scale time series embedding module, and the multi-stage feature fusion module. Therefore, this application designs a residual connection to ensure that the original time series information can be directly passed to the final fusion stage. Specifically, the residual connection module performs a residual connection between the obtained fused embedding and the initially input cultural and tourism historical time series data to obtain the corresponding target embedding. The expression for the residual connection is: Where out is the target embedding. For integration and embedding, X represents historical time-series data on cultural tourism.

[0022] In this embodiment, the target embedding is processed by the output projection layer to predict the corresponding cultural and tourism demand prediction results. The cultural and tourism demand prediction results are a sequence of predicted values ​​for passenger flow and various consumption in the target area over a future period of time.

[0023] This application provides a cultural tourism demand prediction system based on multi-scale data fusion. The system comprises: a data acquisition module for acquiring historical time-series data of cultural tourism in a target area. The historical time-series data is a three-dimensional tensor structure, including batch size, review window length, and number of variables; a multi-resolution processing module for extracting multi-resolution time-frequency features from the historical time-series data in parallel in the time and frequency domains, stitching them together to obtain stitched features, and fusing the stitched features along the time dimension to obtain corresponding fused features; and a multi-scale time-series embedding module for processing the fused features. The system employs a multi-scale patch partitioning method, stacking the partitioning results to obtain multi-scale embeddings. A multi-stage feature fusion module performs inter-variable feature interaction and intra-variable feature exchange processing on the multi-scale embeddings to obtain the corresponding target multi-scale embedding. A feature fusion enhancement module performs weighted fusion of the obtained spliced ​​features, fused features, and target multi-scale embeddings to obtain the corresponding fused embedding. A residual connection module performs residual connection between the fused embedding and the historical time-series data of cultural tourism to obtain the target embedding. An output projection layer processes the target embedding and predicts the corresponding cultural tourism demand prediction results.

[0024] The core of this application lies in providing a cultural tourism demand forecasting system for digital cultural tourism platforms, enabling multi-scale prediction of diverse operational indicator sequence data. The system's core is an end-to-end prediction model based on a Patch Fusion MLP (Multilayer Perceptron) framework, designed to overcome current challenges in processing cultural tourism operational data, such as multi-scale patterns, multivariate variable coupling, and complex time-frequency feature fusion. Its processing flow consists of three main modules: a multi-resolution processing module, a multi-scale time series embedding module, and a multi-stage feature fusion module. For the multi-resolution processing module, this application, for the first time, systematically extracts and effectively fuses multi-resolution features simultaneously in the time domain (trend, seasonality) and frequency domain (periodicity, noise) within an MLP-based framework, addressing the weakness of existing methods in handling complex periodicity and noise. For the multi-scale time series embedding module, by processing patches of different lengths in parallel, it breaks the limitation of single-patch scale in current long-term data prediction, achieving simultaneous capture of multi-granularity time-series patterns of macro trends and micro details in cultural tourism operational data. The multi-stage feature fusion module not only performs the interaction between time steps of the sequence (traditional time-domain modeling), but more importantly, it explicitly models the complex coupling relationship between different operational indicator sequences, which improves the accuracy of the prediction model when processing multivariate sequences such as cultural and tourism consumption data, thereby effectively improving the accuracy of prediction.

[0025] In conjunction with the above embodiments, in one implementation, this application also provides a cultural tourism demand prediction system based on multi-scale data fusion. In this cultural tourism demand prediction system based on multi-scale data fusion, the multi-resolution processing module includes: a time-domain feature extraction module, a frequency-domain feature extraction module, a feature concatenation module, and a time-series multilayer perceptron module. The time-domain feature extraction module is used to perform multi-resolution enhancement and attenuation processing on the input historical time-series data of cultural tourism, respectively, to obtain corresponding enhanced features and attenuated features. The frequency-domain feature extraction module is used to perform Fourier transform on the input historical time-series data of cultural tourism, and then perform sequence zero-padding and Top-K frequency selection processing, respectively, to obtain a first feature after sequence zero-padding and a second feature after Top-K frequency selection. The feature concatenation module is used to concatenate the enhanced features, the attenuated features, the first feature, and the second feature to obtain concatenated features. The time-series multilayer perceptron module is used to fuse the concatenated features in the time dimension to obtain fused features.

[0026] In this embodiment, although time-series data is inherently discrete, its continuous representation masks the potential loss of crucial information between sampling points. Modeling at a single high resolution not only increases computational burden and overfitting risk due to data redundancy but may also make it difficult for the model to capture global macro trends. Therefore, to resolve the contradiction between information loss and data redundancy, this application designs a time-domain feature extraction module within the multi-resolution processing module. This module performs multi-resolution enhancement and reduction operations on the input cultural tourism historical time-series data in the time dimension to obtain the enhanced and reduced features corresponding to the cultural tourism historical time-series data, as expressed below:

[0027]

[0028] in, To enhance features, To reduce the characteristics, This indicates the input historical time series data of cultural tourism.

[0029] In this embodiment, from the perspective of the frequency domain, frequency analysis can capture periodicity and frequency characteristics that are difficult to express explicitly in the time domain, which is crucial for improving the accuracy of multi-resolution information fusion. However, given the discreteness and finiteness of frequency domain sampling points, Fourier transform has inherent limitations, namely, some spectral lines containing key information may be missed. Simultaneously, the original spectrum often contains some irrelevant noise, which causes unnecessary trouble for subsequent processing. To address this issue, this application designs a frequency domain feature extraction module within the multi-resolution processing module. After performing a Fourier transform on the input cultural tourism historical time series data, this module performs zero-padding and Top-K frequency selection operations on the Fourier transform results to obtain the first feature after zero-padding and the second feature after Top-K frequency selection. Through zero-padding, this application effectively increases the sampling point density in the frequency domain, thereby obtaining denser spectral lines, enabling the prediction model of this application to observe and capture previously unseen spectral components. Meanwhile, the Top-K frequency selection strategy selects the most representative frequency components based on the spectral amplitude, effectively preserving the fitted curve while minimizing the influence of noise or unrelated signals. The expressions for sequence zero-padding and Top-K frequency selection are as follows:

[0030]

[0031] in, As the first feature, As the second feature, This indicates the input historical time series data of cultural tourism.

[0032] In this embodiment, the feature stitching module in the multi-resolution processing module is used to stitch together the enhanced features, weakened features, first features, and second features obtained by the time-domain feature extraction module and the frequency-domain feature extraction module to obtain the corresponding stitched features, expressed as:

[0033] in, For splicing features, B represents batch size, N represents feature length, and 2L represents feature temperature.

[0034] In this embodiment, to effectively integrate and utilize the multi-resolution features generated by the temporal feature extraction module and the frequency domain feature extraction module, this application designs a temporal multilayer perceptron module in the multi-resolution processing module. This module adaptively learns and fuses heterogeneous features from the temporal and frequency domains, and enhances the fused features. The temporal multilayer perceptron module fuses the spliced ​​features along the temporal dimension. The fused features contain splicing information from both the frequency and temporal domains, aiming to further remove noise and redundant information, preparing for subsequent multi-scale processing. The temporal multilayer perceptron module compresses and refines the spliced ​​features using a temporal multi-level programming (MLP) to obtain the corresponding fused features, expressed as:

[0035] In this embodiment, through the temporal multilayer perceptron module, the multi-resolution processing module can automatically learn the dependencies between features at different resolutions and generate a unified feature representation (i.e., fused feature) that integrates temporal and frequency domain information. This approach avoids simple feature concatenation or averaging and can more effectively utilize multi-scale information. Figure 2 As shown, Figure 2 The Time Block is the time-domain feature extraction module, and the Fourier Block is the frequency-domain feature extraction module.

[0036] In conjunction with the above embodiments, in one implementation, this application also provides a cultural tourism demand prediction system based on multi-scale data fusion. In this multi-scale data fusion-based cultural tourism demand prediction system, the multi-scale time series embedding module includes: a multi-scale extraction module, a feature compression module, a dimension unification module, a location encoding module, and a feature stacking module. The multi-scale extraction module is used to perform multi-scale patch partitioning processing on the fused features using an overlapping sliding window strategy to obtain a corresponding patch feature matrix, wherein the patch feature matrix consists of a set of patch features at different scales. The feature compression module is used to map the patch features from their original length to a preset compression dimension to obtain the patch compressed features corresponding to each patch feature. The dimension unification module is used to map the patch compressed features to a unified dimension to obtain the corresponding patch unified features. The location encoding module is used to inject location encoding information into the patch unified features at each time step to obtain the corresponding patch encoded features. The feature stacking module is used to combine the patch encoded features at all scales through a stacking operation to obtain the corresponding multi-scale embedding.

[0037] In this application, the feature compression module, the dimension unification module, and the position encoding module are deployed independently for each patch scale.

[0038] In this embodiment, traditional single-scale feature extraction strategies have limitations, failing to simultaneously capture temporal dependencies at different granularities within a time series, resulting in incomplete feature representations. To address this issue, this application designs a multi-scale time series embedding module. This module systematically extracts and integrates feature representations at different temporal granularities by processing patch information from multiple time scales in parallel, thereby providing richer and more comprehensive temporal information for subsequent tasks. Time series data inherently contains complex multi-scale patterns, such as short-term intraday fluctuations, medium-term intraweekly periodic changes, and long-term annual trends. However, traditional single-scale patch methods often struggle to fully capture all these important periodic patterns and subtle local details, leading to an incomplete understanding of the sequence's intrinsic dynamics. Therefore, this application proposes a multi-scale patch feature strategy executed in parallel using the multi-scale time series embedding module designed in this application.

[0039] In this embodiment, the specific execution process of the multi-scale patch feature strategy is as follows: First, the multi-scale extraction module performs multi-scale patch division processing on the obtained fusion features through an overlapping sliding window strategy to obtain the corresponding patch feature matrix. The patch feature matrix consists of a set of patch features at different scales, that is, the patch matrix records the patch features at different scales of the fusion features.

[0040] To facilitate understanding, the process of partitioning a fused feature patch into a patch feature matrix is ​​illustrated here. For example, the process of partitioning a fused feature patch involves dividing it into a set of patch features at scale 'a', a set of patch features at scale 'b', and a set of patch features at scale 'c'. The patch feature matrix corresponding to this fused feature then records the patch features at scales a, b, and c, respectively. In other words, this fused feature is partitioned into patch features at scale a, scale b, and scale c, respectively. Thus, the patch feature matrix corresponding to this fused feature records the patch feature partitioning results at three different scales.

[0041] For example, fusion features rich in time and frequency domain information Segmenting overlapping patch features into M patch scales, the number of patch features for each patch scale. It can be determined by the sequence length at the corresponding patch scale. Patch length The step size S determines the specific calculation expression: To maintain temporal continuity and enhance feature robustness, this application sets the step size S to half the length of the patch features, i.e. .

[0042] In this embodiment, the feature compression module, dimensionality unification module, and positional encoding module in the multi-scale time series embedding module are deployed independently for each patch scale. That is, for each patch scale obtained from the patch scale division, there will be a corresponding feature compression module, a dimensionality unification module, and a positional encoding module to perform corresponding processing operations on the patch features at that patch scale. For example, if patch features are divided into three patch scales, there will be three feature compression modules, three dimensionality unification modules, and three positional encoding modules.

[0043] In this embodiment, the feature compression module is used to map the patch features at its corresponding patch scale from the original length to the compressed dimension. By flattening the two dimensions into one and performing local information aggregation, the compressed patch feature corresponding to the patch feature is obtained. For the i-th patch feature at the M-th patch scale... The process of feature compression can be represented as: B indicates the batch size. This indicates the number of patch features at the corresponding patch scale. Indicates dimension.

[0044] In this embodiment, the patch features are mapped from their original length to compressed dimensions using a feature compression module. After obtaining the corresponding compressed patch features, the dimensionality unification module maps the compressed patch features to a unified dimension to obtain the corresponding unified patch features. This achieves balanced dimensionality distribution, ensuring that patch features at each patch scale have equal status in subsequent processing. The corresponding expression is: The dimension unification module here prepares for subsequent positional encoding by mapping the patch compressed features to the unified dimension required for positional encoding. Specifically, the unified dimension mapped by the dimension unification module is the same for each patch scale.

[0045] In this embodiment, although patch features can capture local temporal patterns, the absolute location information of the original cultural and tourism historical time series data may be partially obscured during the patching process. Therefore, injecting location encoding information into each time step helps the prediction model better understand and utilize the temporal structure. Specifically, the location encoding module injects location encoding information into the patch unified features at each time step to obtain the corresponding patch encoded features. This application adopts the classic sine-cosine location encoding scheme, defined as follows:

[0046] Where pos represents the position index in the input sequence, and i represents the index in the hidden dimension. It is the size of the hidden dimension of the prediction model.

[0047] In this embodiment, in the multi-scale time series embedding module, positional encoding is uniformly applied to the patch feature results at M different patch scales. Specifically, for each patch scale... The positional codes are fused using an addition operation: .

[0048] In this embodiment, after positional encoding enhancement, the patch-encoded features at M patch scales are combined through a stacking operation to form the final multi-scale embedding, as shown in the following expression:

[0049]

[0050] in, The multi-scale embedding is obtained after processing by the multi-scale time series embedding module.

[0051] Compared to traditional splicing operations, stacking operations maintain the independence of features at each scale, providing greater flexibility for subsequent cross-scale feature interactions and fusion. This design enables the prediction model to provide rich multi-scale representation information while preserving the integrity of features at each scale.

[0052] In conjunction with the above embodiments, in one implementation, this application also provides a cultural tourism demand prediction system based on multi-scale data fusion. In this cultural tourism demand prediction system based on multi-scale data fusion, the multi-stage feature fusion module includes: a dimension transformation module, an inter-variable feature interaction module, a reshaping module, and an intra-variable feature exchange module. The dimension transformation module is used to perform dimension transformation on the multi-scale embedding, placing the variable dimension in the multi-scale embedding as the last dimension to obtain a corresponding dimension-transformed multi-scale embedding. The inter-variable feature interaction module is used to perform information exchange and feature enhancement processing on the variable dimension of the dimension-transformed multi-scale embedding to obtain a first interactive multi-scale embedding. The reshaping module is used to reshape the first interactive multi-scale embedding to obtain a reshaped multi-scale embedding. The intra-variable feature exchange module is used to perform intra-variable deep feature transformation processing on the reshaped multi-scale embedding to obtain a target multi-scale embedding.

[0053] In this embodiment, after obtaining multi-scale embeddings, effectively integrating heterogeneous feature information from different processing stages and establishing complex dependencies between variables becomes the key and challenging aspects. To address these two issues, this application proposes a multi-stage feature fusion module based on MLP. This module mainly consists of two key components: an inter-variable feature interaction module and an intra-variable feature exchange module. The inter-variable feature interaction module (Inter-Variable MLP) is specifically responsible for establishing and learning the inter-dependencies between different variables. The intra-variable feature exchange module (Intra-Variable MLP) is specifically responsible for performing deep nonlinear feature transformations and dimensionality mappings within a single variable. Figure 3 As shown, Figure 3 The diagram shows the structure of the feature interaction module between variables; for example... Figure 4 As shown, Figure 4 The diagram shows the structure of the intravariable feature exchange module.

[0054] In this embodiment, after processing by the multi-scale time series embedding module, the prediction model ignores the complex interactions between variables, limiting the improvement of prediction performance. To address this, this application designs an inter-variable feature interaction module to enhance feature representation by explicitly modeling the dependencies and interaction patterns between multiple variables. Specifically, the design goals of this inter-variable feature interaction module include: capturing the correlation and synergistic effects between different variables; learning nonlinear interaction patterns between variables; and providing contextual information from other related variables for each variable, thereby improving the expressive power of univariate features. The multi-scale embedding obtained after processing by the multi-scale time series embedding module is described below. The multi-scale embedding is transformed using the dimension transformation module, so that the variable dimension is located in the last dimension to facilitate subsequent interactions between variables. The corresponding dimension-transformed multi-scale embedding is obtained as follows:

[0055] In this embodiment, for the dimensional transformation multi-scale embedding obtained by the dimensional transformation module, cross-variable feature interaction is established through the processing of the inter-variable feature interaction module. That is, the inter-variable MLP is applied to establish cross-variable feature interaction, and the corresponding first interactive multi-scale embedding is obtained, as shown in the following expression:

[0056]

[0057] in, This indicates the first interaction multi-scale embedding.

[0058] In this embodiment, through the inter-variable feature interaction module, the prediction model can exchange information and enhance features along the variable dimension, so that the representation of each variable not only includes its own temporal pattern, but also incorporates complementary information from related variables. The inter-variable feature interaction module consists of a linear layer, a GELU activation function, and a dropout layer.

[0059] In this embodiment, after obtaining the first interactive multi-scale embedding through the inter-variable feature interaction module in the multi-stage feature fusion module, the intra-variable feature exchange module in the same multi-stage feature fusion module further processes the first interactive multi-scale embedding. Before the intra-variable feature exchange module in the multi-stage feature fusion module further processes the first interactive multi-scale embedding, the first interactive multi-scale embedding is first reshaped by the reshaping module to perform dimensional adjustment and information aggregation, thereby obtaining a reshaped multi-scale embedding. This prepares for subsequent intra-variable deep feature extraction. The reshaping expression is: For the reconstructed multi-scale embedding obtained through reshaping, in-variable deep feature transformation is performed through the intra-variable feature exchange module, that is, in-variable MLP is applied to perform in-variable deep feature transformation to obtain the corresponding target multi-scale embedding, as shown in the following expression:

[0060] in, This represents the multi-scale embedding of the target.

[0061] In this embodiment, after processing by the intra-variable feature exchange module, the prediction model can extract high-order nonlinear features within each variable, mapping the feature dimension to a dimension unified with the time-frequency features, thereby enhancing the expressive power and discriminative power of the features. The intra-variable feature exchange module consists of two linear layers, two GELU activation functions, and a random dropout layer.

[0062] Based on the same inventive concept, this application provides a method for predicting cultural and tourism demand based on multi-scale data fusion, which is applied to the cultural and tourism demand prediction system based on multi-scale data fusion described in the first aspect of this application, such as... Figure 5 As shown, the method includes: Step S01: Obtain the cultural and tourism historical time series data of the target area through the data acquisition module. The cultural and tourism historical time series data is a three-dimensional tensor structure, which includes batch size, review window length, and number of variables. Step S02: Extract multi-resolution time-frequency features from the cultural and tourism historical time series data in parallel in the time and frequency domains using a multi-resolution processing module, and stitch them together to obtain stitched features. Then, fuse the stitched features in the time dimension to obtain corresponding fused features. Step S03: The fused features are divided into multi-scale patches using the multi-scale time series embedding module, and the partitioning results are stacked to obtain multi-scale embeddings; Step S04: The multi-scale embedding is processed by the multi-stage feature fusion module to perform inter-variable feature interaction and intra-variable feature exchange to obtain the corresponding target multi-scale embedding; Step S05: The obtained spliced ​​features, fused features and target multi-scale embedding are weighted and fused by the feature fusion enhancement module to obtain the corresponding fused embedding; Step S06: Perform residual connections between the fused embedding and the original input features using the residual connection module to obtain the target embedding; Step S07: Process the target embedding through the output projection layer to predict the corresponding cultural tourism demand prediction results.

[0063] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily necessary for the embodiments of this application.

[0064] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0065] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0070] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0071] The above provides a detailed description of the cultural tourism demand prediction system and method based on multi-scale data fusion provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A cultural tourism demand forecasting system based on multi-scale data fusion, characterized in that, The system includes: a data acquisition module, a multi-resolution processing module, a multi-scale time series embedding module, a multi-stage feature fusion module, a feature fusion enhancement module, a residual connection module, and an output projection layer; The data acquisition module is used to acquire the cultural and tourism historical time series data of the target area. The cultural and tourism historical time series data is a three-dimensional tensor structure, which includes batch size, review window length and number of variables. The multi-resolution processing module is used to extract multi-resolution time-frequency features from the cultural and tourism historical time series data in parallel in the time domain and frequency domain, and to stitch them together to obtain stitched features, and to fuse the stitched features in the time dimension to obtain corresponding fused features. The multi-scale time series embedding module is used to perform multi-scale patching on the fused features and stack the patching results to obtain multi-scale embedding. The multi-stage feature fusion module is used to perform inter-variable feature interaction and intra-variable feature exchange processing on the multi-scale embedding to obtain the corresponding target multi-scale embedding. The feature fusion enhancement module is used to perform weighted fusion of the obtained spliced ​​features, the fused features and the target multi-scale embedding to obtain the corresponding fused embedding. The residual connection module is used to perform residual connection on the fused embedding and the cultural tourism historical time series data to obtain the target embedding; The output projection layer is used to process the target embedding and predict the corresponding cultural and tourism demand prediction results.

2. The cultural tourism demand prediction system based on multi-scale data fusion according to claim 1, characterized in that, The multi-resolution processing module includes: a time-domain feature extraction module, a frequency-domain feature extraction module, a feature stitching module, and a time-series multilayer perceptron module; The time-domain feature extraction module is used to perform multi-resolution enhancement and reduction processing on the input cultural and tourism historical time series data to obtain corresponding enhancement features and reduction features; The frequency domain feature extraction module is used to perform Fourier transform on the input cultural and tourism historical time series data, and then perform sequence zero-padding and Top-K frequency selection processing to obtain the first feature after sequence zero-padding and the second feature after Top-K frequency selection. The feature splicing module is used to splice the enhanced feature, the weakened feature, the first feature, and the second feature to obtain spliced ​​features; The temporal multilayer perceptron module is used to fuse the spliced ​​features in the time dimension to obtain fused features.

3. The cultural tourism demand prediction system based on multi-scale data fusion according to claim 1, characterized in that, The multi-scale time series embedding module includes: a multi-scale extraction module, a feature compression module, a dimension unification module, a position encoding module, and a feature stacking module; The multi-scale extraction module is used to perform multi-scale patch partitioning on the fused features using an overlapping sliding window strategy to obtain a corresponding patch feature matrix, wherein the patch feature matrix consists of a set of patch features at different scales. The feature compression module is used to map the patch features from their original length to a preset compression dimension to obtain the patch compressed features corresponding to each patch feature. The dimension unification module is used to map the patch compressed features to a unified dimension to obtain the corresponding patch unified features. The location encoding module is used to inject location encoding information into the patch unified features at each time step to obtain the corresponding patch encoded features; The feature stacking module is used to combine patch-encoded features at all scales through stacking operations to obtain the corresponding multi-scale embedding.

4. The cultural tourism demand forecasting system based on multi-scale data fusion according to claim 3, characterized in that, The feature compression module, the dimension unification module, and the position encoding module are deployed independently for each patch scale.

5. The cultural tourism demand prediction system based on multi-scale data fusion according to claim 1, characterized in that, The multi-stage feature fusion module includes: a dimension transformation module, an inter-variable feature interaction module, a reshaping module, and an intra-variable feature exchange module; The dimension transformation module is used to perform dimension transformation on the multi-scale embedding, placing the variable dimension in the multi-scale embedding into the last dimension, and obtaining the corresponding dimension-transformed multi-scale embedding. The variable feature interaction module is used to perform information exchange and feature enhancement processing on the variable dimension of the dimensional transformation multi-scale embedding to obtain the first interactive multi-scale embedding. The reshaping module is used to reshape the first interactive multi-scale embedding to obtain a reshaped multi-scale embedding. The intra-variable feature exchange module is used to perform intra-variable deep feature transformation processing on the reshaped multi-scale embedding to obtain the target multi-scale embedding.

6. The cultural tourism demand forecasting system based on multi-scale data fusion according to claim 5, characterized in that, The feature interaction module between variables consists of a linear layer, a GELU activation function, and a random dropout layer.

7. A cultural tourism demand prediction system based on multi-scale data fusion according to claim 5, characterized in that, The intravariable feature exchange module consists of two linear layers, two GELU activation functions, and a random dropout layer.

8. A cultural tourism demand forecasting system based on multi-scale data fusion according to claim 3, characterized in that, When dividing patch features at a corresponding scale, the multi-scale extraction module sets the division step size to half the length of the patch features at that corresponding scale.

9. A cultural tourism demand forecasting system based on multi-scale data fusion according to claim 1, characterized in that, The historical time series data of the cultural and tourism area shall include at least: multi-dimensional historical time series data of the target area containing visitor flow and various consumption indicators over a period of time; The obtained cultural and tourism demand forecast results are a sequence of predicted values ​​for passenger flow and various consumption items over a future period of time.

10. A method for predicting cultural tourism demand based on multi-scale data fusion, characterized in that, The method applied to the multi-scale data fusion-based cultural tourism demand forecasting system according to any one of claims 1 to 9 includes: The data acquisition module acquires the cultural and tourism historical time series data of the target area. The cultural and tourism historical time series data is a three-dimensional tensor structure, which includes batch size, review window length, and number of variables. The multi-resolution processing module extracts multi-resolution time-frequency features from the historical time series data of cultural tourism in parallel in the time and frequency domains, and splices them to obtain spliced ​​features. Furthermore, the spliced ​​features are fused in the time dimension to obtain corresponding fused features. The fused features are divided into multi-scale patches using a multi-scale time series embedding module, and the partitioning results are stacked to obtain multi-scale embeddings. The multi-stage feature fusion module performs inter-variable feature interaction and intra-variable feature exchange processing on the multi-scale embedding to obtain the corresponding target multi-scale embedding. The obtained spliced ​​features, fused features and target multi-scale embedding are weighted and fused by the feature fusion enhancement module to obtain the corresponding fused embedding. The target embedding is obtained by performing a residual connection between the fused embedding and the original input features through a residual connection module. The target embedding is processed by the output projection layer to predict the corresponding cultural and tourism demand.

Images