Hybrid deep learning automobile sales prediction method fusing multi-source features and application
By combining semi-supervised learning and hybrid deep learning methods, and integrating MLP and AttBiLSTM modules to process multi-source features, the problem of traditional models being unable to fuse time-dependent and non-temporal features is solved, achieving efficient fusion and accurate prediction of automobile sales forecasts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIFANG UNIV OF NATITIES
- Filing Date
- 2025-12-03
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies struggle to effectively integrate time-dependent and non-time-series features, resulting in insufficient feature extraction, incomplete information fusion, and inadequate model learning in subjective evaluation data scenarios, failing to effectively uncover nonlinear interaction relationships between different types of features.
A semi-supervised learning feature clustering method is used to extract features from subjective evaluation data. Static non-time-series data and time-series data are processed by the MLP module and AttBiLSTM module in the parallel feature extraction layer, respectively. Regularization is used to suppress model overfitting, forming a unified high-dimensional feature vector representation. The Adam optimization algorithm is then used to train the car sales prediction model.
It achieves joint extraction and fusion of multi-source features, improves the accuracy and stability of automobile sales prediction, can automatically identify key features under unlabeled conditions, improves the problem of long sequence dependencies, and enhances the model's learning efficiency and prediction accuracy.
Smart Images
Figure CN121961641A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a hybrid deep learning method and application for predicting automobile sales by fusing multi-source features, belonging to the field of hybrid deep learning technology that fuses multi-source features. Background Technology
[0002] With the rapid development of artificial intelligence and big data technologies, the data types involved in practical applications are becoming increasingly complex. This includes the automotive market, where factors influencing car sales often simultaneously include time-dependent features, static attributes, and subjective evaluation-based non-time-series features. Traditional algorithms often struggle to capture the interaction patterns between time-series and non-time-series features when faced with these complex mixed datasets. This leads to insufficient feature extraction, incomplete information fusion, and inadequate model generalization ability, thus affecting the performance of subsequent prediction tasks.
[0003] Defects and shortcomings of existing technology:
[0004] 1. Existing time-series sales models can only effectively handle time-dependent features and cannot take into account non-time-series attribute data; while static sales models lack the ability to learn dynamic time patterns, resulting in poor performance in multimodal data scenarios.
[0005] 2. Most studies model time-series data and non-time-series data separately, without establishing a unified feature representation space. This results in insufficient feature fusion, limited information flow, and an inability to effectively explore the nonlinear interaction relationships between different types of features.
[0006] 3. In subjectively collected data scenarios, some samples lack clear labels, and traditional supervised learning methods cannot make use of a large amount of unlabeled data, resulting in insufficient model learning or biased results. Summary of the Invention
[0007] The purpose of this invention is to solve the problems existing in the prior art, and to provide a hybrid deep learning method and application for predicting automobile sales by fusing multi-source features.
[0008] The objective of this invention is achieved through the following technical solution:
[0009] A hybrid deep learning method for predicting car sales by fusing multi-source features includes the following steps:
[0010] Step 1: Input the subjective evaluation data into the semi-supervised feature generation layer, and extract the features from the subjective evaluation data by the semi-supervised learning feature clustering method;
[0011] Step 2: Input the static non-time series data and the subjective evaluation data from Step 1 into the MLP module in the parallel feature extraction layer, and input the time series data related to automobile sales into the AttBiLSTM module in the parallel feature extraction layer for feature extraction.
[0012] Step 3: The features output from the MLP module and AttBiLSTM module in Step 2 are transformed through progressive feature transformation. At the same time, regularization is introduced to suppress model overfitting and deep fusion is performed to form a unified high-dimensional feature vector representation. The merged feature vector is input into the fusion optimization layer, and the features are gradually refined through nonlinear transformation and dimensionality reduction compression to obtain the car sales prediction model.
[0013] Step 4: Train the car sales prediction model using the Adam optimization algorithm to obtain the loss function of the car sales prediction model that is minimized; evaluate the performance of the car sales prediction model using validation data during training to prevent overfitting; stop training when no further improvement in validation performance is observed; finally, evaluate the performance of the car sales prediction model by iterating through the training and validation of each layer and calculating the validation loss to determine the effectiveness of the final car sales prediction model.
[0014] Preferably, the specific steps of the semi-supervised learning feature clustering method described in step one are as follows:
[0015] First, K-means clustering algorithm is used for unsupervised grouping to discover potential feature patterns:
[0016] (1)
[0017] in, Indicates the first The feature vector of each sample, i.e. the feature input of the subjective evaluation sample; Indicates the first One cluster; Indicates the first The center of a cluster is the mean vector of all samples in that cluster. Indicates the number of clusters in the cluster; The goal of clustering is to make sample points as close as possible to their respective cluster centers.
[0018] Then, pseudo-labels are generated for each sample:
[0019] (2)
[0020] Then, a random forest classifier is trained using pseudo-labels as supervision signals, and the feature importance is calculated using the following formula:
[0021] (3)
[0022] in, Representation of features Importance score; Indicates the number of trees in a random forest; Indicates the first Among the trees, based on characteristics A set of nodes with splitting characteristics; Indicates at node The number of samples divided into segments; Indicates at node Use features The reduction in information gain after partitioning; This represents the total number of samples.
[0023] The parameters of the data alignment and fusion model are calculated based on the feature importance ranking results obtained from the above calculations.
[0024] Preferably, the data alignment and fusion model formula is as follows:
[0025] (4)
[0026] in, For car sales dataset, , , This is a dataset for subjective evaluation. , These correspond to the weights of the two datasets, respectively. , , These results are derived from the semi-supervised feature generation layer.
[0027] Preferably, the MLP module in step two is used for subjective evaluation of high-dimensional feature nonlinear mapping of non-temporal inputs, and the MLP module formula is as follows:
[0028] (5)
[0029] in, They represent the 1st and the 2nd respectively. Passing the exam The activation function of the layer, Representing layers 0 to 1 and 1 respectively To the The weight matrix of the layer, From level 0 to level 1 and level 2 respectively To the Layer bias, This represents the output of the neuron.
[0030] Preferably, the AttBiLSTM module in step two includes: a first-layer BiLSTM, an attention layer, and a second-layer BiLSTM; after the output of the first-layer BiLSTM, the attention layer is introduced, and the attention layer extracts the key patterns in the output of the first-layer BiLSTM, focusing on important time steps and features by dynamically allocating weights. Subsequently, the output of the attention layer is refined and features are extracted by the second-layer BiLSTM.
[0031] Preferably, the specific steps for feature extraction in step two are as follows:
[0032] Step 21: Temporal Feature Extraction from the AttBiLSTM Module
[0033] Step 211: Input the time series data related to car sales, i.e., the original time series. Calculate the forward output at the first layer of BiLSTM. With backward output Splicing bidirectional output , to obtain the sequence ;
[0034] Step 212: At the attention layer Enhance attention mechanisms to generate queries ,key Sum Calculate attention weights and sum them by weight. ;
[0035] Steps 2-13: Input the second layer BiLSTM and take the output of the final time step as the final temporal feature. ;
[0036] Step 22: Non-temporal feature extraction from the MLP module
[0037] Input static non-time series data, i.e., raw non-time series data Output the final non-time series features .
[0038] Preferably, the merged feature vector in step three is:
[0039] Combine the outputs of the AttBiLSTM module and the MLP module. ;
[0040] The fusion optimization layer includes: a first fully connected network and a second fully connected network;
[0041] The first fully connected layer is used for nonlinear mapping and feature representation enhancement of the input features. It contains 128 neurons, uses the SELU activation function to preserve high-order interaction features, and combines regularization methods. Layer normalization is used to accelerate the convergence speed of the car sales prediction model while maintaining gradient stability, and the Dropout rate is introduced to further suppress overfitting. The output then enters the second fully connected layer.
[0042] The second fully connected layer performs dimensionality reduction and compression on the high-dimensional features output by the first fully connected layer, and removes redundant features to obtain a more compact fused feature representation, containing 64 neurons; it also uses the SELU activation function to maintain self-normalization properties.
[0043] (6)
[0044] in, and These are respectively represented as extracted temporal features and non-temporal features such as subjective evaluation categories; and These are weight matrices, used to perform linear transformations on time-series and non-time-series static features, mapping them to the same dimensional space to achieve additive fusion. Indicates the bias term; This represents the activation function.
[0045] Preferably, the loss function of the car sales prediction model described in step four is as follows:
[0046] (7)
[0047] in, Mean square error, The mean absolute error, The regularization coefficient is . and Separate expressions and The weighting coefficients, The weight set includes all weight matrices of AttBiLSTM and MLP. Represents a single weight matrix in the set;
[0048] By dynamically adjusting the weights, an optimal balance between accuracy and robustness is achieved.
[0049] An application of a hybrid deep learning method for predicting car sales based on the fusion of multi-source features in the problem of car sales prediction is presented. Subjective evaluation data is input into a semi-supervised feature generation layer; static non-time-series data including car model parameters, maximum mileage, battery capacity, government subsidies, and the number of charging piles nationwide, along with subjective evaluation data, are input into an MLP module; time-series data including car sales and prices are input into an AttBiLSTM module, fully combining the advantages of the two modules in extracting different features for feature extraction; then, the feature outputs are deeply fused to form a unified high-dimensional feature vector representation, and the merged feature vector is input into a fusion optimization layer.
[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0051] 1. This invention realizes the joint feature extraction and fusion of mixed data of time-series features, static attributes and subjective evaluation non-time-series features, and designs a data alignment fusion model to achieve more efficient fusion of automobile sales information, avoid information redundancy and overfitting caused by splicing, and optimize the accuracy of automobile sales prediction.
[0052] 2. This invention can automatically identify key features for predicting car sales even without labels, thereby improving the learning efficiency of car sales prediction models.
[0053] 3. This invention significantly improves the ability to identify key time steps by introducing an attention mechanism, thereby improving the problem of long sequence dependencies.
[0054] 4. The weighted loss function and L2 regularization term in this invention improve the stability and noise resistance of the automobile sales prediction model. Attached Figure Description
[0055] Figure 1 This is a schematic diagram of the structure of the hybrid deep learning method for predicting automobile sales by integrating multiple source features, and its application, as described in this invention. Detailed Implementation
[0056] The present invention will be further described in detail below with reference to the accompanying drawings: This embodiment is implemented under the premise of the technical solution of the present invention, and detailed implementation methods are given, but the protection scope of the present invention is not limited to the following embodiments.
[0057] like Figure 1 As shown, this invention proposes a hybrid deep learning method for predicting car sales by fusing multi-source features, based on the optimization of sales prediction accuracy. It is a feature extraction algorithm based on a fusion structure of semi-supervised learning and AttBiLSTM-MLP, used for car sales datasets that simultaneously contain time-dependent, static attribute-based, and subjective evaluation types. The algorithm consists of three layers: a semi-supervised feature generation layer, a parallel feature extraction layer, and a fusion optimization layer.
[0058] (1) Semi-supervised feature generation layer
[0059] The primary purpose of this layer is to address subjective evaluation data. Considering that this type of data is difficult to directly correlate with sales, the semi-supervised feature generation layer proposes a method based on semi-supervised learning feature clustering to extract features from subjective evaluation data. First, the K-means clustering algorithm is used for unsupervised grouping to discover potential feature patterns.
[0060] (1)
[0061] in, Indicates the first The feature vector of each sample, i.e. the feature input of the subjective evaluation sample; Indicates the first One cluster; Indicates the first The center of a cluster is the mean vector of all samples in that cluster. Indicates the number of clusters in the cluster; The goal of clustering is to make sample points as close as possible to their respective cluster centers.
[0062] Then, pseudo-labels are generated for each sample:
[0063] (2)
[0064] Then, a random forest classifier is trained using pseudo-labels as supervision signals, and the feature importance is calculated using the following formula:
[0065] (3)
[0066] in, Representation of features Importance score; Indicates the number of trees in a random forest; Indicates the first Among the trees, based on characteristics A set of nodes with splitting characteristics; Indicates at node The number of samples divided into segments; Indicates at node Use features The reduction in information gain after partitioning; This represents the total number of samples.
[0067] Calculate the parameters of the data alignment and fusion model based on the feature importance ranking results.
[0068] (2) Parallel feature extraction layer
[0069] The parallel feature extraction layer is a key component of hybrid deep learning prediction methods that fuse multi-source features. This layer aims to leverage the model's strengths in extracting different features by hierarchically modeling the fused dataset. Static non-time-series data, including vehicle model parameters, maximum mileage, battery capacity, government subsidies, and the number of charging stations nationwide, as well as subjective evaluation data, are input into the MLP module. Time-series data, including vehicle sales and prices, are input into the AttBiLSTM module, fully combining the advantages of both modules in extracting different features.
[0070] The AttBiLSTM structure is as follows:
[0071] This architecture is designed as a two-layer bidirectional LSTM layer with an attention mechanism introduced in the middle layer. An attention layer is introduced after the output of the first BiLSTM layer to extract key patterns from the output of the first BiLSTM layer. By dynamically allocating weights, it focuses on important time steps and features, effectively solving the problem of long sequence dependencies and further enhancing the model's ability to capture information from key time periods in time series data. Subsequently, another BiLSTM layer is used to refine and extract features.
[0072] Temporal feature extraction using the AttBiLSTM module, the steps are as follows:
[0073] Input time series data related to car sales, i.e., the original time series. Calculate the forward output at the first layer of BiLSTM. With backward output Splicing bidirectional output , to obtain the sequence ;
[0074] Attention layer Enhance attention mechanisms to generate queries ,key Sum Calculate attention weights and sum them by weight. ;
[0075] Input the second layer BiLSTM, and take the output at the final time step as the final temporal feature. ;
[0076] The non-temporal feature extraction steps of the MLP module are as follows:
[0077] Input static non-time series data, i.e., raw non-time series data Output the final non-time series features .
[0078] The MLP structure is as follows:
[0079] (4)
[0080] in, They represent the 1st and the 2nd respectively. Passing the exam The activation function of the layer, Representing layers 0 to 1 and 1 respectively To the The weight matrix of the layer, From level 0 to level 1 and level 2 respectively To the Layer bias, This represents the output of the neuron.
[0081] MLP structures are used for high-dimensional feature nonlinear mapping of non-temporal inputs such as subjective evaluation.
[0082] (3) Fusion Optimization Layer
[0083] This layer employs a method of progressively extracting shared features in a nonlinear mapping, preserving the characteristic differences between different vehicle models. Regularization effectively suppresses noise interference, improving model stability. The feature outputs of the AttBiLSTM and MLP models are deeply fused to form a unified high-dimensional feature representation. The merged feature vector is input into the first fully connected layer, which contains 128 neurons. The SELU activation function is used to preserve high-order interaction features. Layer normalization maintains gradient stability while accelerating model convergence, and Dropout is introduced to further suppress overfitting. The output then enters the second fully connected layer, containing 64 neurons, which also uses the SELU activation function to maintain self-normalization. The second fully connected layer removes redundancy and noise from the data through dimensionality reduction compression, ensuring the model can extract more refined features. This achieves more efficient information fusion, avoiding information redundancy and overfitting caused by splicing.
[0084] (5)
[0085] in, and These are respectively represented as extracted temporal features and non-temporal features such as subjective evaluation categories; and These are weight matrices, used to perform linear transformations on time-series and non-time-series static features, mapping them to the same dimensional space to achieve additive fusion. Indicates the bias term; This represents the activation function.
[0086] The model's loss function is designed as a multi-objective weighted form:
[0087] (6)
[0088] in, Mean square error, The mean absolute error, The regularization coefficient is . and Separate expressions and The weighting coefficients, The weight set includes all weight matrices of AttBiLSTM and MLP. Represents a single weight matrix in the set;
[0089] By dynamically adjusting the weights, an optimal balance between accuracy and robustness is achieved.
[0090] Data alignment and fusion model
[0091] (7)
[0092] in For BYD's car sales data set, , , This is a dataset for subjective evaluation. , The weights corresponding to the two datasets are respectively used in this study. , , These results are derived from the semi-supervised feature generation layer.
[0093] The specific algorithm process of this invention is as follows:
[0094] After time-series data is input into the AttBiLSTM layer, it undergoes two forward and backward feature extraction processes. Through a two-layer structure, an attention mechanism is introduced in the middle layer, allowing the car sales prediction model to more deeply understand and learn the temporal characteristics in the time-series data, further filtering and preserving key information. Non-time-series static data and subjective evaluation data are input into the MLP layer, where dense layers learn the non-linear relationships between the data and capture information from them.
[0095] After feature extraction through AttBiLSTM and MLP layers, the extracted hidden outputs are fused through a dual fully connected layer. The dual fully connected layer integrates time-series features extracted from the dataset with non-time-series static features through progressive feature transformation and regularization to suppress overfitting, avoiding information fragmentation. Simultaneously, it refines features progressively through nonlinear transformation and dimensionality reduction compression, improving the accuracy and generalization ability of the car sales prediction model. The merged feature vector is then input into the first fully connected layer, which contains 128 neurons. The SELU activation function is used to preserve high-order interaction features, and layer normalization maintains gradient stability while accelerating model convergence. Dropout is introduced to further suppress overfitting. The output then enters the second fully connected layer, containing 64 neurons, which also uses the SELU activation function to maintain self-normalization. The second fully connected layer removes redundancy and noise from the data through dimensionality reduction compression, ensuring that the car sales prediction model can extract more refined features. This structure employs a method of progressively extracting shared features in a nonlinear mapping, preserving the differences between different features, and effectively suppressing noise interference through a regularization mechanism to improve the stability of the car sales forecasting model.
[0096] The Adam optimization algorithm was used to train the car sales forecasting model to minimize the loss function. To ensure a balance between generalization and training efficiency, validation data was used to evaluate the model's performance during training to prevent overfitting. Training was stopped when no further improvement in validation performance was observed. Finally, the model's performance was evaluated by iterating through the training and validation of each layer and calculating the validation loss to determine the effectiveness of the final car sales forecasting model.
[0097] Example 1
[0098] Step 1: Feature Processing and Encoding
[0099] 1> Categorical variable encoding, semi-supervised parameter solving (From a method based on semi-supervised learning feature clustering)
[0100] 2> Data Standardization Processing
[0101] Step 2: Parallel Feature Processing
[0102] 1>AttBiLSTM Temporal Feature Extraction
[0103] Input the original time series Calculate the forward output at the first layer of BiLSTM. With backward output Splicing bidirectional output , to obtain the sequence
[0104] right Enhance attention mechanisms to generate queries ,key Sum Calculate attention weights and sum them by weight.
[0105] Input the second layer BiLSTM, and take the output at the final time step as the final temporal feature.
[0106] 2>MLP Non-Temporal Feature Extraction
[0107] Input raw non-time series Output the final non-time series features
[0108] Step 3: Feature Fusion
[0109] 1> Concatenate the AttBiLSTM and MLP outputs.
[0110] 2> Fusion via dual fully connected layers
[0111] The first fully connected layer: used for non-linear mapping and feature representation enhancement of the input features, contains 128 neurons, uses the SELU activation function, and combines regularization methods.
[0112] The second fully connected layer is used to reduce the dimensionality of the high-dimensional features output by the first fully connected layer and remove redundant features, resulting in a more compact fused feature representation. It contains 64 neurons.
[0113] Step 4: Training and Optimization
[0114] 1> Minimize the loss function using the Adam optimizer
[0115] 2> Set up an early stopping mechanism to prevent overfitting and save the best-performing model on the validation set.
[0116] Step 5: Predict the output.
[0117] The above description is merely a preferred embodiment of the present invention. These specific embodiments are different implementations based on the overall concept of the present invention, and the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A hybrid deep learning method for predicting car sales by fusing multi-source features, characterized in that, Includes the following steps: Step 1: Input the subjective evaluation data into the semi-supervised feature generation layer, and extract the features from the subjective evaluation data by the semi-supervised learning feature clustering method; Step 2: Input the static non-time series data and the subjective evaluation data from Step 1 into the MLP module in the parallel feature extraction layer, and input the time series data related to automobile sales into the AttBiLSTM module in the parallel feature extraction layer for feature extraction. Step 3: The features output from the MLP module and AttBiLSTM module in Step 2 are transformed through progressive feature transformation. At the same time, regularization is introduced to suppress model overfitting and deep fusion is performed to form a unified high-dimensional feature vector representation. The merged feature vector is input into the fusion optimization layer, and the features are gradually refined through nonlinear transformation and dimensionality reduction compression to obtain the car sales prediction model. Step 4: Train the car sales prediction model using the Adam optimization algorithm to obtain the loss function of the car sales prediction model that is minimized; evaluate the performance of the car sales prediction model using validation data during training to prevent overfitting; stop training when no further improvement in validation performance is observed; finally, evaluate the performance of the car sales prediction model by iterating through the training and validation of each layer and calculating the validation loss to determine the effectiveness of the final car sales prediction model.
2. The hybrid deep learning method for predicting car sales by fusing multi-source features according to claim 1, characterized in that, The specific steps of the semi-supervised learning feature clustering method described in step one are as follows: First, K-means clustering algorithm is used for unsupervised grouping to discover potential feature patterns: （1） in, Indicates the first The feature vector of each sample, i.e. the feature input of the subjective evaluation sample; Indicates the first One cluster; Indicates the first The center of a cluster is the mean vector of all samples in that cluster. Indicates the number of clusters in the cluster; The goal of clustering is to make sample points as close as possible to their respective cluster centers. Then, pseudo-labels are generated for each sample: （2） Then, a random forest classifier is trained using pseudo-labels as supervision signals, and the feature importance is calculated using the following formula: （3） in, Representation of features Importance score; Indicates the number of trees in a random forest; Indicates the first Among the trees, based on characteristics A set of nodes with splitting characteristics; Indicates at node The number of samples divided into segments; Indicates at node Use features The reduction in information gain after partitioning; This represents the total number of samples. The parameters of the data alignment and fusion model are calculated based on the feature importance ranking results obtained from the above calculations.
3. The hybrid deep learning method for predicting car sales by fusing multi-source features according to claim 2, characterized in that, The formula for the data alignment and fusion model is as follows: （4） in, For car sales dataset, , , This is a dataset for subjective evaluation. , These correspond to the weights of the two datasets, respectively. , , These results are derived from the semi-supervised feature generation layer.
4. The hybrid deep learning method for predicting car sales by fusing multi-source features according to claim 1, characterized in that, The MLP module mentioned in step two is used for subjective evaluation of high-dimensional feature nonlinear mappings of non-temporal inputs. The MLP module formula is as follows: （5） in, They represent the 1st and the 2nd respectively. Passing the exam The activation function of the layer, Representing layers 0 to 1 and 1 respectively To the The weight matrix of the layer, From level 0 to level 1 and level 2 respectively To the Layer bias, This represents the output of the neuron.
5. The hybrid deep learning method for predicting car sales by fusing multi-source features according to claim 1, characterized in that, The AttBiLSTM module described in step two includes: a first-layer BiLSTM, an attention layer, and a second-layer BiLSTM. The output of the first-layer BiLSTM is introduced into the attention layer, which extracts key patterns from the output of the first-layer BiLSTM and focuses on important time steps and features by dynamically allocating weights. Subsequently, the output of the attention layer is refined and features are extracted by the second-layer BiLSTM.
6. The hybrid deep learning method for predicting car sales by fusing multi-source features according to claim 5, characterized in that, The specific steps for feature extraction described in step two are as follows: Step 21: Temporal Feature Extraction from the AttBiLSTM Module Step 211: Input the time series data related to car sales, i.e., the original time series. Calculate the forward output at the first layer of BiLSTM. With backward output Splicing bidirectional output , to obtain the sequence ; Step 212: At the attention layer Enhance attention mechanisms to generate queries ,key Sum Calculate attention weights and sum them by weight. ; Steps 2-13: Input the second layer BiLSTM and take the output of the final time step as the final temporal feature. ; Step 22: Non-temporal feature extraction from the MLP module Input static non-time series data, i.e., raw non-time series data Output the final non-time series features .
7. The hybrid deep learning method for predicting car sales by fusing multi-source features according to claim 6, characterized in that, The merged feature vector described in step three is: Combine the outputs of the AttBiLSTM module and the MLP module. ; The fusion optimization layer includes: a first fully connected network and a second fully connected network; The first fully connected layer is used for nonlinear mapping and feature representation enhancement of the input features. It contains 128 neurons, uses the SELU activation function to preserve high-order interaction features, and combines regularization methods. Layer normalization is used to accelerate the convergence speed of the car sales prediction model while maintaining gradient stability, and the Dropout rate is introduced to further suppress overfitting. The output then enters the second fully connected layer. The second fully connected layer performs dimensionality reduction and compression on the high-dimensional features output by the first fully connected layer, and removes redundant features to obtain a more compact fused feature representation, containing 64 neurons; it also uses the SELU activation function to maintain self-normalization properties. （6） in, and These are respectively represented as extracted temporal features and non-temporal features such as subjective evaluation categories; and These are weight matrices, used to perform linear transformations on time-series and non-time-series static features, mapping them to the same dimensional space to achieve additive fusion. Indicates the bias term; This represents the activation function.
8. The hybrid deep learning method for predicting car sales by fusing multi-source features according to claim 1, characterized in that, The loss function of the car sales prediction model described in step four is as follows: （7） in, Mean square error, The mean absolute error, The regularization coefficient is . and Separate expressions and The weighting coefficients, The weight set includes all weight matrices of AttBiLSTM and MLP. Represents a single weight matrix in the set; By dynamically adjusting the weights, an optimal balance between accuracy and robustness is achieved.
9. An application of a hybrid deep learning method for predicting automobile sales based on the fusion of multi-source features according to any one of claims 1-8 in the problem of automobile sales prediction, characterized in that, Input subjective evaluation data into the semi-supervised feature generation layer; The system will input static, non-time-series data, including vehicle model parameters, maximum vehicle mileage, battery capacity, government subsidies, and the number of charging piles nationwide, as well as subjective evaluation data, into the MLP module. Time-series data, including vehicle sales and prices, will be input into the AttBiLSTM module. The system will fully combine the advantages of the two modules in extracting different features to perform feature extraction. Then, the feature outputs will be deeply fused to form a unified high-dimensional feature vector representation. The merged feature vector will be input into the fusion optimization layer.