Traffic flow prediction method, system and device based on neural component decomposition and dynamic adaptive fusion

Abstract

The present application belongs to the technical field of intelligent transportation system, and discloses a traffic flow prediction method, system and device based on neural component decomposition and dynamic adaptive fusion. The present application comprises: performing reversible instance normalization on input multi-element traffic data, and adaptively decomposing the normalized data into trend, seasonality and residual components through a neural decomposition module; extracting time features for each channel by using an independent recurrent neural network, and generating each component through a shared channel mixing network; obtaining the prediction values of each component, dynamically weighting the prediction values, and obtaining a combined prediction tensor; performing the inverse process of reversible instance normalization on the combined prediction tensor to generate the final prediction result about future traffic conditions. The present application can effectively reduce the learning difficulty of the prediction model, capture the nonlinear time-varying relationship between each component, and significantly improve the accuracy and reliability of traffic flow prediction, thereby providing technical support for efficient traffic relief and signal light control.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation systems technology, specifically relating to a traffic flow prediction method, system, and device based on neural component decomposition and dynamic adaptive fusion. Background Technology

[0002] Intelligent Transportation Systems (ITS) are a key technological approach to solving traffic congestion, and accurate traffic flow prediction is the core and prerequisite for achieving efficient traffic management and control (such as dynamic signal timing, ramp control, and travel route planning).

[0003] Urban traffic flow data is a typical multivariate time series data, collected by numerous sensors deployed on the road network (such as geomagnetic coils and microwave detectors). It exhibits high nonlinearity, dynamic time-varying characteristics, and complex spatiotemporal correlations. Traditional statistical models, such as ARIMA (Autoregressive Integrated Moving Average), are difficult to effectively model complex traffic flow patterns such as peak hours and holiday effects due to their linear assumptions. In recent years, deep learning models, represented by recurrent neural networks (RNNs) and the Transformer architecture, have become the mainstream technology in traffic flow prediction due to their powerful nonlinear modeling capabilities.

[0004] To analyze complex traffic flow sequences with greater precision, time series decomposition is an effective strategy. This strategy aims to deconstruct raw traffic data into several more easily understood and modeled basic components. Specifically, the raw sequence can be viewed as a superposition of three core components: first, the trend component, which reflects the long-term, slow direction of traffic flow changes caused by macroeconomic factors such as urban planning changes, population growth, or the opening of new routes; second, the seasonal component, which captures traffic patterns with strong periodicity, such as morning and evening rush hours within a day, weekday and weekend tidal differences within a week, and even travel patterns on specific holidays throughout the year; and third, the residual component, which represents traffic flow fluctuations caused by sudden, random events such as traffic accidents, severe weather, or temporary traffic control that are difficult to explain using the first two models.

[0005] However, existing technologies still face two major challenges when combining decomposition ideas with deep learning models for traffic flow prediction:

[0006] 1. Rigidity and fragmentation of decomposition mechanisms: Traditional decomposition methods (such as STL) typically employ fixed moving average filters, which cannot adaptively capture the ever-changing traffic patterns across different road segments and time periods. More importantly, the decomposition process is usually treated as a preprocessing step independent of the prediction model, preventing the prediction model from learning during training how to optimally separate the trends and seasonal features most relevant to the prediction target. This fragmented optimization process, especially in end-to-end deep models, limits the overall performance of the model.

[0007] 2. Simplicity and Limitations of Reconstruction Fusion Methods: After predicting each traffic component separately, existing methods often use simple linear addition to synthesize the final result. This approach ignores the complex dynamic interactions between the components. For example, the impact of a traffic accident (residual) occurring during off-peak hours versus during the morning rush hour (seasonal) is drastically different and highly nonlinear on traffic flow. Simple addition cannot capture this "congestion superposition effect," thus severely limiting the model's prediction accuracy under complex traffic conditions.

[0008] Furthermore, traffic flow data is inherently non-stationary and time-varying. Sensors deployed on different road sections and in different areas may collect data with drastically different statistical characteristics (such as mean and variance); even the same sensor may experience data distribution shifts before and after different seasons or special events (such as road construction). This "distribution shift" phenomenon makes it difficult for a model trained on one dataset to generalize to another dataset with a different distribution, severely limiting the model's practical deployment capability and robustness.

[0009] Therefore, designing a traffic flow decomposition mechanism that can effectively overcome data distribution bias, perform end-to-end joint optimization with the prediction model, and be adaptively learnable, coupled with an intelligent fusion strategy that can dynamically model trends, periodic peaks, and nonlinear interactions between sudden events, has become an urgent technical problem to be solved in order to improve the predictive ability, robustness, and reliability of modern intelligent transportation systems. Summary of the Invention

[0010] The purpose of this invention is to address the shortcomings of existing technologies by providing a traffic flow prediction method, system, and device based on neural component decomposition and dynamic adaptive fusion. This method first eliminates the distribution bias of heterogeneous traffic data through reversible normalization technology, and then achieves integrated decomposition and prediction and dynamic reconstruction fusion of complex traffic flow data through an end-to-end learnable framework, thereby significantly improving the accuracy of traffic state prediction and the robustness of the model in dealing with traffic emergencies.

[0011] Specifically, the present invention is implemented using the following technical solutions.

[0012] On one hand, this invention provides a traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion for predicting traffic flow data with a length dimension of H, including:

[0013] Preprocessing: Receive input historical multi-dimensional traffic flow data, where the data has dimensions B×S×C, where B is the number of samples in a batch, S is the length of the time series data, and C is the number of channels. Perform reversible instance normalization on the historical multi-dimensional traffic flow data to obtain normalized input data. ;

[0014] Decomposition Prediction: The normalized input data is decomposed and predicted using a learnable neural decomposition prediction module. Decompose and predict into trend prediction components Seasonal forecast components and residual prediction components ;

[0015] Dynamic adaptive fusion weights: For each channel c, c=1,...,C, the trend prediction component is dynamically calculated. Seasonal forecast components and residual prediction components The fusion weights are then combined and summed to form the final prediction result for channel c. The final prediction result combining all channels. This yields a combined prediction tensor; the inverse process of the reversible instance normalization is then performed on the combined prediction tensor to restore it to the original data scale, serving as the final prediction result for future traffic conditions. ;

[0016] The normalized input data Decompose and predict into trend prediction components Seasonal forecast components and residual prediction components The steps include:

[0017] 2-1) Channel Independent Stage: Extract features containing temporal context information independently for each channel and generate the original trend tensor. ;

[0018] 2-2) Channel blending stage: Utilizing the original trend tensor Synthetic trend prediction components ; using the normalized input data and the trend prediction component Synthesize the seasonal forecast components Based on the additive decomposition relationship, the residual prediction components are obtained through the following algebraic operations. :

[0019] ;

[0020] in, The normalized input data Dimension alignment is performed to match the traffic flow data representation with the predicted length dimension H.

[0021] Furthermore, the process involves independently extracting features containing temporal context information for each channel and generating the original trend tensor. include:

[0022] 2-1-1) For the normalized input data Each of the C channels Use a dedicated long short-term memory network encoder for this channel. To process the time series data of this channel ,in, This involves extracting data from all B samples in the current batch, specifically the data from the c-th channel across all time steps, where the dimension of all time steps is S, to generate a high-dimensional temporal feature embedding. ;

[0023]

[0024] Where B is the number of samples contained in a batch, and S is the length of the time series data. It is the dimension of the hidden layer of the Long Short-Term Memory network. This means extracting the data from all B samples and the c-th channel of all time steps in the current batch being processed, where the dimension of all time steps is S;

[0025] 2-1-2) Embed the high-dimensional temporal features of each channel through a linear shrinkage layer shared by all channels. Compressed into univariate feature sequences And through a nonlinear activation function The linear contraction layer consists of a weight vector. and bias terms definition:

[0026] ;

[0027] 2-1-3) The univariate feature sequences of all channels Stacking along the channel dimension forms the final original trend tensor. , .

[0028] Furthermore, the utilization of the original trend tensor Synthetic trend prediction components include:

[0029] The channel vector of the original trend tensor Z at each time step t is obtained by using a time-step shared channel hybrid multilayer perceptron. A nonlinear transformation is performed to capture the instantaneous dependencies between different channels and generate a trend vector at that moment. Where b and t are standard indices, representing the b-th sample and t-th time step in the batch, respectively, b=1,...,B. It is the slice vector of the original trend tensor Z at the index position. This slice vector is formed by extracting features for each channel in the channel-independent stage and then stacking along the channel dimension as described in step 2-1-3). Its C dimensions correspond to the state of the C channels at that moment.

[0030] Furthermore, the use of the normalized input data and the trend prediction component Synthesize the seasonal forecast components include:

[0031] From the normalized input data Subtract the generated trend prediction component from the middle To obtain the detrended sequence ;

[0032] The detrended sequence is processed using another independent channel-mixing multilayer perceptron containing at least one hidden layer. The process is performed to generate the seasonal forecast component. .

[0033] Furthermore, for each channel c, c=1,...,C, the trend prediction component is dynamically calculated. Seasonal forecast components and residual prediction components The final fusion weights are then combined and summed to form the final prediction result for channel c. include;

[0034] 3-1) The trend prediction component of channel c Seasonal forecast components and residual prediction components Element-wise addition forms a context feature vector. ;

[0035] 3-2) Use a residual gating mechanism to adjust the static baseline weights. Dynamic weight adjustment item The final fusion weights are calculated by adding them together. ,

[0036]

[0037] Wherein, the static benchmark weight It is a learnable parameter vector containing three values, corresponding to the global base weights of the trend, seasonality, and residual components, respectively. The static baseline weights... Shared across all channels and all samples; the dynamic weight adjustment term It is a dynamic adjustment term for the current sample and the current channel, with... The dimension, where B is the batch size, H is the set future length, and 3 corresponds to three components; the dynamic weight adjustment term By using the context feature vector Input to a shared gated network And generated;

[0038] 3-3) Use the final fusion weights. For the trend prediction component Seasonal forecast components and residual prediction components By performing a weighted summation, the final prediction result for channel c is obtained. :

[0039]

[0040] Where ⊙ represents element-wise multiplication; For trend prediction components, For seasonal forecast components, For residual prediction components; The final fusion weights for channel c and its corresponding component k; This is the predicted value of the corresponding component k for channel c.

[0041] Furthermore, the static benchmark weights Dynamic weight adjustment item Addition includes: taking the static benchmark weights The data is copied along the batch dimension B and the prediction length dimension H, expanding its dimensions to B×H×3 to match the dimensions of the dynamic weight adjustment term, and then summed.

[0042] On the other hand, the present invention also provides a traffic flow prediction system based on neural component decomposition and dynamic adaptive fusion, which realizes the above-mentioned traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion. The traffic flow prediction system based on neural component decomposition and dynamic adaptive fusion includes a neural decomposition prediction module and a dynamic adaptive fusion module.

[0043] The neural decomposition prediction module receives historical multivariate traffic flow data from various traffic sensors in the urban road network and performs reversible instance normalization processing on it to obtain normalized input data to eliminate distribution bias; the normalized input data is decomposed and predicted into trend prediction components, seasonal prediction components and residual prediction components.

[0044] The dynamic adaptive fusion module dynamically calculates the fusion weights of the trend prediction component, seasonal prediction component, and residual prediction component for each channel and performs a weighted summation to fuse them into the final prediction result for that channel c. It then combines the final prediction results of all channels to obtain a combined prediction tensor. Finally, it performs the inverse process of the reversible instance normalization on the combined prediction tensor to restore it to the original data scale, which serves as the final prediction result for future traffic conditions.

[0045] In another aspect, the present invention provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion described above.

[0046] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion described above.

[0047] In another aspect, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion described above.

[0048] The beneficial effects of the traffic flow prediction method, system, and device based on neural component decomposition and dynamic adaptive fusion of the present invention are as follows:

[0049] 1. Improve traffic prediction accuracy and robustness: By decomposing the complex original sequence into smoother and more regular subsequences, the learning difficulty of downstream models is reduced; and the nonlinear superposition effect between regular traffic patterns and emergencies is effectively captured through a dynamic fusion mechanism, which together improves the accuracy of the final prediction results and the reliability under complex road conditions.

[0050] 2. Enhanced modeling capabilities for complex traffic phenomena: The learnable decomposition and prediction modules and dynamic fusion modules enable the entire prediction framework to be optimized end-to-end. The model can adaptively learn the decomposition and reconstruction strategies best suited to the current road network state, thereby accurately modeling diverse and complex traffic phenomena such as weekday morning rush hours, holiday travel tidal waves, and even traffic accidents.

[0051] 3. Task decoupling and efficiency improvement: The traffic flow prediction method, system and device based on neural component decomposition and dynamic adaptive fusion of the present invention removes the difficult task of separating non-orthogonal time components in traffic flow (such as separating random fluctuations from peak patterns) from the core prediction model, so that it can focus on modeling a "cleaner" signal, thereby improving the model's learning efficiency and performance.

[0052] 4. Improved interpretability: The traffic flow prediction method, system and device based on neural component decomposition and dynamic adaptive fusion of the present invention explicitly decomposes components with clear traffic physical meaning such as trends and seasonality. Furthermore, the dynamic weights generated by the dynamic adaptive fusion module also reveal the degree of dependence of the model on each component under different traffic conditions, providing a decision window for traffic managers to understand the prediction results and formulate targeted traffic management strategies (such as determining whether congestion is caused by periodic peaks or sudden accidents).

[0053] 5. Enhanced model generalization ability and deployment feasibility: By introducing a reversible instance normalization step before decomposition, this invention effectively overcomes the problem of data distribution offset between different sensors and different time periods, so that the prediction model does not need to be trained separately for each road segment. This significantly enhances the model's ability to process non-stationary data and its generalization performance, and reduces the difficulty of actual deployment. Attached Figure Description

[0054] Figure 1 This is a structural block diagram of the main functional modules of the system according to an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the embodiments and the accompanying drawings.

[0057] One embodiment of the present invention is a traffic flow prediction system based on neural component decomposition and dynamic adaptive fusion, employing a "channel-independent to channel-hybrid" (CI→CM) architecture, such as... Figure 1 As shown, it includes a neural decomposition prediction module and a dynamic adaptive fusion module.

[0058] The neural decomposition prediction module receives historical multivariate traffic flow data from multiple traffic sensors in the urban road network and performs reversible instance normalization to eliminate distribution bias. The data after reversible instance normalization is then decomposed and predicted into at least one trend prediction component, one seasonal prediction component, and one residual prediction component. First, independent time features are extracted from each channel of the input historical multivariate time series. Then, the trend and seasonal components are generated through a channel mixing mechanism. The residual prediction component is obtained by subtracting the trend and seasonal prediction components from the input historical multivariate time series. The trend prediction component reflects long-term traffic flow changes brought about by urban development; the seasonal prediction component reflects periodic patterns such as morning and evening peak hours; and the residual prediction component reflects sudden events such as traffic accidents.

[0059] The dynamic adaptive fusion module dynamically weights the trend prediction component, seasonal prediction component, and residual prediction component to generate a final prediction result about future traffic conditions. The dynamic adaptive fusion module generates a set of dynamic fusion weights independently for each channel based on the prediction value of each component.

[0060] The traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion of the present invention is used to predict traffic flow data for a set future length H (H is a natural number), such as... Figure 2 As shown, it includes the following steps:

[0061] I. Pretreatment

[0062] Receive input historical multi-dimensional traffic flow data (For example, traffic flow and speed sequences from multiple sensors in the urban road network). The historical multi-dimensional traffic flow data has a dimension of B×S×C, where B is the batch size, S is the historical sequence length, and C is the number of channels, which in this embodiment is the number of traffic sensors. The input historical multi-dimensional traffic flow data is subjected to Reversible Instance Normalization (ReVIN) processing to obtain normalized input data. This step aims to eliminate distributional differences between different sequences and stabilize model training. The decomposition and fusion in this invention are both performed on the normalized data scale.

[0063] II. Decomposition and Prediction

[0064] A learnable neural decomposition prediction module is used to analyze normalized historical multivariate traffic flow data. Decompose and predict into trend prediction components Seasonal forecast components and residual prediction components The specific structure and implementation of the learnable neural decomposition prediction module are jointly defined by the channel-independent phase (2-1) and the channel-mixed phase (2-2) described below.

[0065] This step aims to process the received normalized input data. Directly outputs the predicted values ​​of the future trend component, seasonal component, and residual component for future time steps (prediction length H): Trend prediction component Seasonal forecast components and residual prediction components .

[0066] Normalized input data Decompose and predict into trend prediction components Seasonal forecast components and residual prediction components The steps include:

[0067] 2-1) Channel-independent stage: Temporal feature extraction

[0068] Features containing temporal context information are extracted independently for each channel (i.e., the monitoring sequence of each traffic sensor), and an original trend tensor of intermediate states is generated. The specific method is as follows:

[0069] 2-1-1) For the normalized input data Each of the C channels It uses a channel-specific Long Short-Term Memory (LSTM) encoder. To process the time series data of this channel Generate a high-dimensional temporal feature embedding :

[0070]

[0071] Where B is the batch size and S is the length of the time series data. It is the dimension of the hidden layer of the Long Short-Term Memory network. This indicates that the data from the c-th channel of all samples (dimension B) and all time steps (dimension S) in the batch is extracted.

[0072] 2-1-2) High-dimensional features of each channel are represented by a linear shrinkage layer shared by all channels. Compressed into a single variable feature sequence And through a nonlinear activation function (e.g., ReLU or GELU). The linear shrinkage layer consists of a weight vector. and bias terms definition.

[0073]

[0074] 2-1-3) Univariate feature sequences of all channels Stacking along the channel dimension forms an intermediate state of the original trend tensor. , .

[0075] 2-2) Channel blending stage: Utilizing the original trend tensor Synthetic trend prediction components Seasonal forecast components and residual prediction components Specifically, this includes:

[0076] 2-2-1) Composite Trend Prediction Component :

[0077] A hybrid multilayer perceptron (MLP) with shared time steps is denoted as... The original trend tensor Z is processed. This MLP operates independently at each time step t, processing the channel vector at that time. A nonlinear transformation is performed, where b and t are standard indices, representing the b-th sample and t-th time step in the batch, respectively, and b=1,...,B. This is a slice of the original trend tensor Z at that index position. This slice vector is formed by extracting features for each channel in the channel independence stage described in step 2-1), followed by stacking along the channel dimensions as described in step 2-1-3). Its C dimensions correspond to the states of the C channels at that moment. This nonlinear transformation is used to capture the instantaneous dependencies between different channels (i.e., modeling the spatial correlation between different road segments), thereby generating the trend vector at that moment. :

[0078]

[0079] in, t is a channel hybrid multilayer perceptron (MLP) that shares weights across all time steps t; b is the batch index; T is an identifier indicating that this MLP is specifically designed to handle trend components.

[0080] 2-2-2) Synthetic seasonal forecast components :

[0081] From the normalized input data described in step one Subtract the generated trend prediction component from the middle To obtain the detrended sequence Using another independent channel-mixed multilayer perceptron containing at least one hidden layer, denoted as... For detrended sequences Processing is performed to generate seasonal forecast components. This predictive component primarily captures periodic patterns in traffic flow, such as morning and evening peak hours.

[0082]

[0083] in, This represents the seasonal prediction component vector generated at time step t for the b-th sample. It is a channel-mixed MLP used to synthesize seasonal forecast components; It is a detrended sequence The feature vector of the b-th sample at time step t.

[0084] 2-2-3) Determine the residual prediction components :

[0085] Based on the additive decomposition relationship, the residual prediction components can be directly obtained through algebraic operations. This component mainly reflects fluctuations caused by random events such as traffic accidents.

[0086]

[0087] in, For normalized input data Perform dimension alignment to match the representation after the predicted length H.

[0088] III. Dynamic Adaptive Fusion Weights

[0089] For each channel c, the trend prediction component is dynamically calculated. Seasonal forecast components and residual prediction components The fusion weights (prediction length H) are then weighted and summed to form the final prediction result for channel c. This fusion process is also channel-independent. The final prediction result combines all channels. (For example, stacking along the channel dimension) yields a combined prediction tensor. Perform the inverse process of the reversible instance normalization on the combined prediction tensor to restore it to the original data scale, which serves as the final prediction result for future traffic conditions. .

[0090] For any channel c, the specific steps of the dynamic adaptive fusion process are as follows:

[0091] 3-1) Generate contextual features

[0092] The trend prediction component of channel c Seasonal forecast components and residual prediction components Add them point by point to form a context feature vector. This vector summarizes the overall signal characteristics of the channel in the future and can be used to judge the overall traffic situation of the road segment in the future (e.g., whether it is smooth or congested).

[0093]

[0094] 3-2) Calculate the dynamic weights

[0095] The final fusion weights are calculated using a residual gating mechanism. The mechanism consists of two parts: static baseline weights and dynamic weight adjustment terms. The final fusion weight is the sum of the two parts.

[0096] Among them, static benchmark weights It is a learnable parameter vector containing three values, corresponding to the global base weights of the trend component, seasonal component, and residual component, respectively. The static baseline weights are shared across all channels and all samples, representing the prior knowledge learned by the model about the general importance of each component.

[0097] Dynamic weight adjustment item It is a dynamic adjustment term for the current sample and the current channel, which is determined by the context feature vector. Input to a shared gated network This mechanism allows the model to adjust weights based on predicted future conditions. For example, when contextual features indicate the possibility of severe congestion caused by both accidents (residuals) and peak periods (seasonality), the model can dynamically increase the weights of these two components.

[0098]

[0099] When adding the static baseline weights and the dynamic weight adjustment term, the dynamic weight adjustment term... have The dimension is defined as follows: B is the batch size, H is the set future length, and 3 corresponds to three components.

[0100] Final fusion weights as follows:

[0101]

[0102] Preferably, in another embodiment, the static benchmark weights With dimension 3, the static baseline weights The data is copied along the batch dimension B and the prediction length dimension H, expanding its dimensions to B×H×3 to match the dimensions of the dynamically adjusted items, and then added together.

[0103] 3-3) Dynamic Adaptive Fusion

[0104] Use the calculated final fusion weights For each component's predicted value (trend prediction component) Seasonal forecast components and residual prediction components We perform a weighted summation to obtain the final prediction result for channel c. :

[0105]

[0106] Where ⊙ represents element-wise multiplication; Indicate the component type; T is the trend forecast component, S is the seasonal forecast component, and R is the residual forecast component. The final fusion weights for channel c and its corresponding component k; This is the predicted value of the corresponding component k for channel c.

[0107] This invention presents a traffic flow prediction method, system, and device based on neural component decomposition and dynamic adaptive fusion, proposing an end-to-end learnable prediction framework. First, reversible instance normalization is performed on multivariate traffic data acquired from road network sensors to effectively overcome data distribution offset issues. Second, a neural decomposition module adaptively decomposes the normalized data into trend, seasonal, and residual components. This module adopts a "channel-independent to channel-mixed" architecture, first using independent recurrent neural networks to extract temporal features for each channel, and then generating each component through a shared channel-mixing network. Subsequently, after obtaining the predicted values ​​of each component, dynamic reconstruction is performed through a dynamic adaptive fusion module. This module employs a residual gating mechanism, combining a set of learnable static baseline weights with an adjustment term dynamically generated based on the prediction context to calculate the final fusion weights, and then performing a weighted summation of the predicted values ​​of each component. The traffic flow prediction method, system, and device based on neural component decomposition and dynamic adaptive fusion of the present invention achieve end-to-end optimization of time series through effective normalization preprocessing, learnable decomposition, and dynamic fusion. This effectively reduces the learning difficulty of the prediction model and can capture the nonlinear time-varying relationship between components, thereby significantly improving the accuracy, robustness, and reliability of traffic flow prediction and providing technical support for achieving efficient traffic management and traffic light control.

[0108] In some embodiments, certain aspects of the above-described techniques may be implemented by one or more processors of a processing system executing software. The software includes one or more sets of executable instructions stored or otherwise tangibly implemented on a non-transitory computer-readable storage medium. The software may include instructions and certain data that, when executed by one or more processors, manipulate one or more processors to perform one or more aspects of the above-described techniques. The non-transitory computer-readable storage medium may include, for example, magnetic or optical disk storage devices, solid-state storage devices such as flash memory, cache, random access memory (RAM), or other non-volatile memory devices. The executable instructions stored on the non-transitory computer-readable storage medium may be source code, assembly language code, object code, or other instruction formats interpreted or otherwise executed by one or more processors.

[0109] Computer-readable storage media can include any storage medium or combination of storage media that can be accessed by a computer system during use to provide instructions and / or data to the computer system. Such storage media can include, but are not limited to, optical media (e.g., optical discs (CDs), digital versatile optical discs (DVDs), Blu-ray discs), magnetic media (e.g., floppy disks, magnetic tapes, or magnetic hard disks), volatile memory (e.g., random access memory (RAM) or cache), non-volatile memory (e.g., read-only memory (ROM) or flash memory), or microelectromechanical systems (MEMS) based storage media. Computer-readable storage media can be embedded in a computing system (e.g., system RAM or ROM), fixedly attached to a computing system (e.g., a magnetic hard disk drive), removably attached to a computing system (e.g., an optical disc or universal serial bus-based (USB) flash memory), or coupled to a computer system via a wired or wireless network (e.g., network-accessible storage (NAS)).

Claims

1. A traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion, used to predict traffic flow data with length dimension H, characterized in that, include: Preprocessing: Receive input historical multi-dimensional traffic flow data, where the data has dimensions B×S×C, where B is the number of samples in a batch, S is the length of the time series data, and C is the number of channels. Perform reversible instance normalization on the historical multi-dimensional traffic flow data to obtain normalized input data. ; Decomposition Prediction: The normalized input data is decomposed and predicted using a learnable neural decomposition prediction module. Decompose and predict into trend prediction components Seasonal forecast components and residual prediction components ; Dynamic adaptive fusion weights: For each channel c, c=1,...,C, the trend prediction component is dynamically calculated. Seasonal forecast components and residual prediction components The fusion weights are then combined and summed to form the final prediction result for channel c. ; The final prediction result combining all channels This yields a combined prediction tensor; the inverse process of the reversible instance normalization is then performed on the combined prediction tensor to restore it to the original data scale, serving as the final prediction result for future traffic conditions. ; The normalized input data Decompose and predict into trend prediction components Seasonal forecast components and residual prediction components The steps include: 2-1) Channel Independent Stage: Extract features containing temporal context information independently for each channel and generate the original trend tensor. ; 2-2) Channel blending stage: Utilizing the original trend tensor Synthetic trend prediction components ; using the normalized input data and the trend prediction component Synthesize the seasonal forecast components Based on the additive decomposition relationship, the residual prediction components are obtained through the following algebraic operations. : ； in, The normalized input data Perform dimension alignment to match the traffic flow data representation with the predicted length dimension H; For each channel c, c=1,...,C, the trend prediction component is dynamically calculated. Seasonal forecast components and residual prediction components The final fusion weights are then combined and summed to form the final prediction result for channel c. include; 3-1) The trend prediction component of channel c Seasonal forecast components and residual prediction components Element-wise addition forms a context feature vector. ; 3-2) Use a residual gating mechanism to adjust the static baseline weights. Dynamic weight adjustment item The final fusion weights are calculated by adding them together. , ， Wherein, the static benchmark weight It is a learnable parameter vector containing three values, corresponding to the global base weights of the trend, seasonality, and residual components, respectively, and the static baseline weights. Shared across all channels and all samples; the dynamic weight adjustment term It is a dynamic adjustment term for the current sample and the current channel, with... The dimension, where B is the batch size, H is the set future length, and 3 corresponds to three components; the dynamic weight adjustment term By using the context feature vector Input to a shared gated network And generated; 3-3) Use the final fusion weights. For the trend prediction component Seasonal forecast components and residual prediction components By performing a weighted summation, the final prediction result for channel c is obtained. : ， Where ⊙ represents element-wise multiplication; For trend prediction components, For seasonal forecast components, For residual prediction components; The final fusion weights for channel c and its corresponding component k; This is the predicted value of the corresponding component k for channel c.

2. The traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion according to claim 1, characterized in that, The process involves independently extracting features containing temporal context information for each channel and generating the original trend tensor. include: 2-1-1) For the normalized input data Each of the C channels Use a dedicated long short-term memory network encoder for this channel. To process the time series data of this channel ,in, This involves extracting data from all B samples in the current batch, specifically the data from the c-th channel across all time steps, where the dimension of all time steps is S, to generate a high-dimensional temporal feature embedding. ; ， Where B is the number of samples contained in a batch, and S is the length of the time series data. It is the dimension of the hidden layer of the Long Short-Term Memory network. This means extracting the data from all B samples and the c-th channel of all time steps in the current batch being processed, where the dimension of all time steps is S; 2-1-2) Embed the high-dimensional temporal features of each channel through a linear shrinkage layer shared by all channels. Compressed into univariate feature sequences And through a nonlinear activation function The linear contraction layer consists of a weight vector. and bias terms definition: ； 2-1-3) The univariate feature sequences of all channels Stacking along the channel dimension forms the final original trend tensor. , .

3. The traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion according to claim 2, characterized in that, The use of the original trend tensor Synthetic trend prediction components include: The channel vector of the original trend tensor Z at each time step t is obtained by using a time-step shared channel hybrid multilayer perceptron. A nonlinear transformation is performed to capture the instantaneous dependencies between different channels and generate a trend vector at that moment. Where b and t are standard indices, representing the b-th sample and t-th time step in the batch, respectively, b=1,...,B. It is the slice vector of the original trend tensor Z at the index position. This slice vector is formed by extracting features for each channel in the channel-independent stage and then stacking them along the channel dimension as described in step 2-1-3). Its C dimensions correspond to the state of the C channels at that moment.

4. The traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion according to claim 1, characterized in that, The normalized input data is used and the trend prediction component Synthesize the seasonal forecast components include: From the normalized input data Subtract the generated trend prediction component from the middle To obtain the detrended sequence ; The detrended sequence is processed using another independent channel-mixing multilayer perceptron containing at least one hidden layer. The process is performed to generate the seasonal forecast component. .

5. The traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion according to claim 1, characterized in that, The static benchmark weight Dynamic weight adjustment item Addition includes: taking the static benchmark weights The data is copied along the batch dimension B and the prediction length dimension H, expanding its dimensions to B×H×3 to match the dimensions of the dynamic weight adjustment term, and then summed.

6. A traffic flow prediction system based on neural component decomposition and dynamic adaptive fusion, implementing the traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion as described in any one of claims 1-5, characterized in that, Includes a neural decomposition prediction module and a dynamic adaptive fusion module; The neural decomposition prediction module receives historical multivariate traffic flow data from various traffic sensors in the urban road network and performs reversible instance normalization processing on it to obtain normalized input data in order to eliminate distribution bias. The normalized input data is decomposed and predicted into trend prediction components, seasonality prediction components, and residual prediction components. The dynamic adaptive fusion module dynamically calculates the fusion weights of the trend prediction component, seasonal prediction component, and residual prediction component for each channel and performs a weighted summation to fuse them into the final prediction result for that channel c. It then combines the final prediction results of all channels to obtain a combined prediction tensor. Finally, it performs the inverse process of the reversible instance normalization on the combined prediction tensor to restore it to the original data scale, which serves as the final prediction result for future traffic conditions.

7. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of the traffic flow prediction method based on neural component decomposition and dynamic adaptive fusion as described in any one of claims 1-5.

Images