New energy automobile production demand prediction method based on multi-modal fusion

Abstract

The invention provides a new energy automobile production demand prediction method based on multi-modal fusion, and relates to the field of production demand prediction. According to the method, firstly, multi-source and multi-mode data are acquired to comprehensively reflect influence factors of new energy automobile production requirements; secondly, on the feature mining level, on the basis of the obtained multi-source and multi-modal data, original fusion features including numerical features, a modal feature set, a total word vector set and an emotion feature set are constructed, then a multiplication crossover method is adopted for processing, and fusion features with the most production demand predictability are fully mined; thirdly, obtaining a corresponding output result by using the pre-trained N-BEATS network and the TFT network, so as to improve the prediction interpretability; and finally, output results of the two networks are further processed by adopting a fusion and moving average method technology, so that the accuracy and the stability of new energy automobile production demand prediction are guaranteed. In addition, the personalized production suggestions provided by the invention are beneficial to reducing the production cost and improving the production efficiency of manufacturers.

Description

Technical Field

[0001] The present invention relates to the field of production demand forecasting, and in particular to a new energy vehicle production demand forecasting method based on multimodal fusion. Background Art

[0002] As a key force in green mobility and industrial transformation and upgrading, new energy vehicles (NEVs) hold strategic significance for promoting sustainable development, reducing environmental pollution, and improving energy efficiency. Accurately forecasting NEV production demand plays a crucial role in ensuring cost-effectiveness and environmental sustainability for NEV manufacturers. The accuracy of production demand forecasts directly impacts manufacturers' ability to effectively plan production, inventory, and procurement strategies, reducing unnecessary costs and improving overall supply chain efficiency and responsiveness.

[0003] Traditional forecasting methods, such as statistical methods like exponential smoothing and generalized autoregressive conditional heteroskedasticity, have some predictive power, but these models often make strong linear assumptions and are limited to specific forecasting scenarios. To address this issue, AI-based machine learning forecasting methods have emerged, such as random forests, Bayesian neural networks, and extreme gradient boosting. These machine learning models are more adept at handling nonlinear relationships and can capture more data patterns and interactions when faced with complex data. However, these methods still have limitations when faced with large datasets and complex feature extraction. In contrast, deep learning, with its deep network structure, can automatically learn multi-level feature representations from massive amounts of data. It can gradually extract features from the data from low-level to high-level levels, eliminating the need for extensive manual feature engineering required by traditional machine learning methods, such as gated recurrent units (GRUs), long short-term memory networks (LSTMs), and temporal convolutional networks (TCNs). By automatically learning complex patterns in the data, it improves forecasting accuracy.

[0004] However, existing deep learning methods have difficulty in effectively integrating rich information from multi-source data, resulting in inaccurate prediction results. Summary of the Invention

[0005] (1) Technical problems solved In response to the shortcomings of the existing technology, the present invention provides a new energy vehicle production demand forecasting method based on multimodal fusion, which solves the technical problem of difficulty in effectively integrating rich information from multi-source data.

[0006] (2) Technical solution To achieve the above objectives, the present invention is implemented through the following technical solutions: A new energy vehicle production demand forecasting method based on multimodal fusion, comprising: Acquire multi-source multi-modal data; the multi-source multi-modal data includes a set of numerical indicators and a textual data set of new energy vehicles, wherein the textual data set includes a news text set, social media comments, and industry reports; Based on the numerical indicator set, respectively obtaining a numerical feature set and a modal feature set; based on the news text set, social media comments and industry reports, obtaining a total word vector set; and based on the news text set, obtaining a sentiment feature set; splicing the numerical features, the modal feature set, the total word vector set, and the sentiment feature set to obtain a first fused feature; and processing the first fused feature using a multiplication cross method to obtain a second fused feature; The second fusion feature is used as the input of the pre-trained N-BEATS network and the TFT network respectively to obtain the corresponding output results; The output results of the two networks are fused, and the sliding average method is used to optimize the fusion results to obtain the final new energy vehicle production demand forecast results.

[0007] Preferably, the new energy vehicle production demand forecasting method further includes: The final new energy vehicle production demand forecast results are used as input to the large language model to obtain personalized production recommendations.

[0008] Preferably, obtaining numerical features based on the numerical indicator set includes: The missing values ​​in the numerical indicator set are processed by the mean filling method, and the abnormal values ​​in the numerical indicator set are processed by the truncation or replacement method of the box plot analysis to obtain complete numerical indicator data; Perform Min-Max normalization processing on the complete numerical indicator data to obtain the numerical features.

[0009] Preferably, obtaining a modal feature set based on the numerical indicator set includes: Performing a Spearman correlation analysis on the current indicator and other indicators in the numerical indicator set to determine the indicator with the highest correlation; Add several groups of different white noises to the indicator with the highest correlation to obtain a set of noisy signals; Perform empirical mode decomposition on each noisy signal to obtain the corresponding intrinsic mode function components; The corresponding intrinsic mode function components are averaged respectively to obtain the collective empirical mode decomposition results as the corresponding modal features; All indicators in the numerical indicator set are traversed to obtain the modal feature set.

[0010] Preferably, obtaining a total word vector set based on the news text set, social media comments, and industry reports includes: Perform text cleaning and standardization on the news text collection, social media comments, and industry reports; We extract a preset number of words from the standardized news text collection, social media comments, and industry reports, and use the FinBERT model to obtain the corresponding word embeddings to obtain a set of news word vectors, a set of social media comment word vectors, and a set of industry report word vectors. The news word vector set, social media comment word vector set, and industry report word vector set are merged, and the principal component analysis method is used to reduce the dimension of the merging results to obtain the total word vector set.

[0011] Preferably, obtaining a set of emotional features based on the news text set includes: Performing text cleaning and standardization on the news text collection; The SnowNLP model is used to convert each standardized news text set into a sentiment score. After traversing all news text sets, the sentiment feature set is obtained.

[0012] Preferably, the adopting the multiplication cross method to process the first fusion feature to obtain the second fusion feature includes: For the numerical features and modal feature sets, numerical features and total word vector sets, numerical features and sentiment feature sets, modal feature sets and sentiment feature sets, and total word vector sets and sentiment feature sets in the first fusion features, the multiplication cross method is used to obtain the corresponding interaction features, and all the interaction features are spliced ​​to obtain the second fusion features.

[0013] A new energy vehicle production demand forecasting system based on multimodal fusion, including: A data acquisition module for acquiring multi-source multi-modal data; the multi-source multi-modal data includes a set of numerical indicators of new energy vehicles and a text-based dataset, wherein the text-based dataset includes a set of news texts, social media comments, and industry reports; A data preprocessing module is configured to obtain a numerical feature set and a modal feature set based on the numerical indicator set; obtain a total word vector set based on the news text set, social media comments, and industry reports; and obtain a sentiment feature set based on the news text set; A feature fusion module is configured to concatenate the numerical features, the modal feature set, the total word vector set, and the sentiment feature set to obtain a first fused feature; and to process the first fused feature using a multiplication cross method to obtain a second fused feature; A result prediction module is used to use the second fusion feature as the input of the pre-trained N-BEATS network and the TFT network respectively to obtain the corresponding output results; The result optimization module is used to fuse the output results of the two networks and optimize the fusion results using the sliding average method to obtain the final new energy vehicle production demand forecast results.

[0014] Preferably, the new energy vehicle production demand forecasting system further includes: The recommendation acquisition module is used to use the final new energy vehicle production demand forecast results as input to the large language model to obtain personalized production recommendations.

[0015] A storage medium stores a computer program for predicting the production demand of new energy vehicles based on multimodal fusion, wherein the computer program enables a computer to execute the new energy vehicle production demand prediction method as described above.

[0016] An electronic device, comprising: One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the new energy vehicle production demand forecasting method as described above.

[0017] (3) Beneficial effects This invention provides a new energy vehicle production demand forecasting method based on multimodal fusion. Compared with the existing technology, it has the following advantages: In the present invention, first, multi-source and multi-modal data are acquired to comprehensively reflect the influencing factors of new energy vehicle production demand; secondly, at the feature mining level, based on the acquired multi-source and multi-modal data, original fusion features including numerical features, modal feature sets, total word vector sets and sentiment feature sets are constructed, and then the multiplication cross method is used for processing to fully mine the fusion features with the most predictive value for production demand; thirdly, the pre-trained N-BEATS network and TFT network are used to obtain the corresponding output results to improve the predictive interpretability; finally, the fusion and sliding average method is used to further process the output results of the two networks to ensure the accuracy and stability of the new energy vehicle production demand forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 A flowchart of a method for predicting production demand of new energy vehicles based on multimodal fusion provided by an embodiment of the present invention; Figure 2 A flowchart of another method for predicting production demand of new energy vehicles based on multimodal fusion provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] The embodiment of the present application solves the technical problem of difficulty in effectively integrating rich information in multi-source data by providing a new energy vehicle production demand forecasting method based on multimodal fusion.

[0022] The technical solution in the embodiments of the present application is to solve the above technical problems, and the overall idea is as follows: The production demand for new energy vehicles is influenced by multiple, heterogeneous factors, including policies, market sentiment, and technological advancements. Existing deep learning methods struggle to effectively integrate the rich information from these multiple data sources. The present invention addresses this issue by making technical improvements in data source acquisition, feature engineering, model selection, and secondary processing of prediction results, aiming to enhance comprehensive data analysis capabilities, optimize feature extraction efficiency, and deeply mine feature information from a multi-dimensional perspective. This improves model algorithm capabilities to output personalized solutions, ultimately achieving high-precision predictions of the production demand for new energy vehicles.

[0023] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0024] Example 1: An embodiment of the present invention provides a method for predicting production demand of new energy vehicles based on multimodal fusion, comprising: S1. Acquire multi-source multi-modal data; the multi-source multi-modal data includes a set of numerical indicators and a textual data set of new energy vehicles, wherein the textual data set includes a news text set, social media comments, and industry reports; S2. Based on the numerical indicator set, respectively obtain a numerical feature set and a modal feature set; based on the news text set, social media comments, and industry reports, obtain a total word vector set; and based on the news text set, obtain a sentiment feature set; S3, concatenating the numerical features, the modal feature set, the total word vector set, and the sentiment feature set to obtain a first fused feature; and processing the first fused feature using a multiplication cross method to obtain a second fused feature; S4. Using the second fusion features as inputs of the pre-trained N-BEATS network and the TFT network, respectively, to obtain corresponding output results; S5. Fuse the output results of the two networks and use the sliding average method to optimize the fusion results to obtain the final new energy vehicle production demand forecast results.

[0025] Through the above solution, the embodiment of the present invention can accurately predict the production demand of new energy vehicles.

[0026] In an optional embodiment, if Figure 2 As shown, the new energy vehicle production demand forecasting method provided by the embodiment of the present invention further includes: S6. Use the final new energy vehicle production demand forecast results as input to the large language model to obtain personalized production recommendations.

[0027] Through the above-mentioned further improved scheme, the embodiment of the present invention helps new energy vehicle manufacturers to scientifically and rationally arrange production plans, optimize inventory management, effectively reduce production costs and improve production efficiency, thereby enhancing the competitiveness of manufacturers in the market.

[0028] The following sections describe the various steps of the above solution in detail: In step S1, multi-source multi-modal data is obtained; the multi-source multi-modal data includes a set of numerical indicators and a text data set of new energy vehicles, and the text data set includes a news text set, social media comments and industry reports.

[0029] For example, the numerical indicator set is defined as .in, represents the nth index, , N represents the total number of indicators, and ,in Represents the characteristic parameter of the mth time of the nth indicator, , M represents the length of time.

[0030] It should be noted that the specific content of the numerical data includes but is not limited to the following aspects: Directly related numerical indicators of new energy vehicles, such as the average transaction price of main-selling models, the average production volume of main-selling models, the average annual mileage, the average sales volume of main-selling models, the transaction price of used cars, automobile import and export volume, automobile production, and automobile demand.

[0031] Raw material-related indicators, such as raw material prices (such as lithium, cobalt, nickel, graphite, aluminum, copper, etc.), rare earth metal prices (such as neodymium, praseodymium, samarium), raw material demand, raw material import and export volume, cobalt mine output, and rare earth metal mining costs.

[0032] Indicators related to upstream and downstream products, such as the number of charging piles built and the capacity utilization rate of automobile manufacturers.

[0033] Fuel-related indicators, such as battery cost, charging cost, battery energy density (such as lithium batteries, solid-state batteries), battery secondary utilization rate, and lithium battery production volume Indirect factors, such as macroeconomic indicators, oil price fluctuations, new energy vehicle subsidy policy amounts, and tax incentives.

[0034] Exemplarily, the text data set obtained in this step includes: News text collection .in, represents the news text at time m, and ,in express The cth news word in the , c∈[1,C], C represents The total number of news words.

[0035] social media comments .in, represents the social media comments at time m, and ,in express The ath comment word in a∈[1,A], A represents The total number of words in social media comments.

[0036] Industry Report .in, represents the industry report text data at time m, and ,in, express The bth industry report word, b∈[1,B], B represents Total number of industry report words.

[0037] In step S2, based on the numerical indicator set, a numerical feature set and a modal feature set are respectively obtained; based on the news text set, social media comments and industry reports, a total word vector set is obtained; and based on the news text set, an emotional feature set is obtained.

[0038] For example, based on a set of numerical indicators , obtain the numerical feature O, including: S201, using a mean filling method to process missing values ​​in the numerical indicator set, and using a truncation or replacement method of a box plot analysis to process abnormal values ​​in the numerical indicator set, to obtain complete numerical indicator data; S202: Perform Min-Max normalization processing on the complete numerical indicator data to obtain the numerical features: in, is the numerical feature corresponding to the nth numerical indicator, and It is the maximum and minimum value of any complete numerical data x.

[0039] For example, based on a set of numerical indicators , obtain the modal feature set R, including: S201′, performing Spearman correlation analysis on the current indicator and other indicators in the numerical indicator set to determine the indicator with the highest correlation.

[0040] S202' adds several groups of different white noises to the indicator with the highest correlation to obtain a noisy signal set.

[0041] S203 ′ performs empirical mode decomposition on each noisy signal to obtain corresponding intrinsic mode function (IMF) components.

[0042] S204′ calculates the mean of the corresponding eigenmode function components respectively to obtain a multi-dimensional ensemble empirical mode decomposition (MEEMD) result as the corresponding modal feature.

[0043] S205' traverses all the indicators in the numerical indicator set to obtain the modal feature set.

[0044] Exemplarily, based on the news text collection, social media comments, and industry reports, a total word vector set is obtained, including: S210: Perform text cleaning and standardization on the news text collection, social media comments, and industry reports.

[0045] S211. Extract a preset number (e.g., D) of words from the standardized news text collection, social media comments, and industry reports, respectively, and use the FinBERT model to obtain corresponding word embeddings to obtain a news word vector set, a social media comment word vector set, and an industry report word vector set.

[0046] Among them, the news word vector set of the mth time is defined , the word vector set of social media comments at time m And the industry report word vector set at time m , thus obtaining the news word vector set , social media comment word vector collection and industry report word vector collection .in, represents the d-th news word vector, represents the d-th social media comment word vector, Represents the word vector of the d-th industry report, where d∈[1,D] and D represents the length of the word vector.

[0047] It should be noted that the FinBERT model is a pre-trained language model based on the BERT (Bidirectional Encoder Representation from Transformers) architecture, specifically designed for text processing tasks in the financial field, such as sentiment analysis, text classification, and entity recognition. This embodiment of the present invention uses this model to perform word embedding operations.

[0048] S213. Merge the news word vector set, the social media comment word vector set, and the industry report word vector set, and use principal component analysis to reduce the dimensionality of the merged results to obtain the total word vector set.

[0049] Among them, the news word vector set A1, social media comment word vector set A2 and industry report word vector set A3 are transposed and spliced ​​to obtain the original total word vector set , where the superscript T represents transpose.

[0050] Afterwards, the principal component analysis method is used to reduce the dimension of the original total word vector set A to obtain the total word vector set after dimension reduction ,in, Represents the word vector feature of the e-th dimension in the original total word vector set A, e∈[1,E], and E represents the total dimension of the word vector feature.

[0051] As you can understand, the present embodiment performs independent word embedding processing on multi-source text data, generating high-dimensional vectors. Because these high-dimensional vectors have varying dimensions after concatenation, hindering subsequent feature extraction, this embodiment employs principal component analysis for dimensionality reduction to optimize the balance between text features and other types of features. This dimensionality reduction step not only effectively reduces the data dimensionality but also facilitates the integration of different text data.

[0052] Exemplarily, obtaining a set of sentiment features based on the news text set includes: S211', performing text cleaning and standardization on the news text collection.

[0053] S212', use the SnowNLP model to convert each standardized news text set into a sentiment score, and after traversing all news text sets, obtain the sentiment feature set Among them, the output range of the sentiment score is [0,1], 0 represents completely negative, 1 represents completely positive, z∈[1,Z], and Z represents the dimension of the sentiment feature.

[0054] It can be understood that the purpose of introducing emotional features in the embodiment of the present invention is to help the model better understand and grasp the emotional information about new energy vehicles in different news texts by quantifying the emotional value of news texts, help identify potential market risks or opportunities, and better predict production demand.

[0055] In step S3, the numerical features, modal feature set, total word vector set and sentiment feature set are concatenated to obtain a first fusion feature; and the first fusion feature is processed by a multiplication cross method to obtain a second fusion feature.

[0056] In this step, the numerical feature O, modal feature index R, and total word vector set are first , and the emotional feature set Q are spliced ​​to form a preliminary first fusion feature B containing multimodal information, which is expressed as: in, Represents a splicing operation.

[0057] Considering that the features in the first fusion feature are arranged independently, this step further uses the multiplication cross method to obtain the second fusion feature, which explicitly strengthens the interaction between features, explores the potential correlation between features, improves the feature discrimination, and introduces nonlinear modeling capabilities to adapt to downstream task requirements.

[0058] It should be noted that the above-mentioned multiplication cross method is also called feature combination, which means multiplying two or more features to achieve nonlinear transformation of the sample space to increase the nonlinear ability of the model.

[0059] The embodiment of the present invention introduces the multiplication cross method to perform interactive processing between different features in parallel in the first fusion feature. For example, this step is for the numerical feature O and the modal feature set R in the first fusion feature B, the numerical feature O and the total word vector set , numerical feature O and sentiment feature set Q, modal feature set R and sentiment feature set Q, total word vector set With the emotional feature set Q, the corresponding interaction features are obtained by multiplication cross method, and all the interaction features are spliced ​​to obtain the second fusion feature .

[0060] To expand on this: (1) The numerical feature O is multiplied and dot-producted with the modal feature set R, so that the model can capture more levels of regularity. It not only focuses on static numerical information, but also reflects the complex dynamics that change over time. Specifically, combining battery life with long-term trends can reveal the impact of battery performance in long-term trends. Conversely, crossing production costs with short-term demand fluctuations may reveal the sensitivity of short-term production cost fluctuations to demand.

[0061] (2) Numerical features O (such as features related to mileage, battery capacity, etc.) and the total word vector set Performing dot product multiplication on features (such as those related to user reviews and industry trend reports) can help subsequent models understand the relationship between technical parameters and consumer perceptions, emotions, or needs. For example, if user reviews repeatedly mention the advantage of "long battery life," and the corresponding numerical feature has a high range, the specific advantages of a particular new energy vehicle can be determined, thereby inferring consumer purchase intentions and product satisfaction.

[0062] (3) The numerical feature O is multiplied and dot-producted with the sentiment feature set Q to reveal the deep connection between technical indicators and user experience. Specifically, if a car has a large battery capacity but there are many negative emotions in user reviews (e.g., although the battery life is long, the charging is too slow), the manufacturer can consider improving the charging speed in subsequent versions; (4) The modal feature set R is multiplied and cross-linked with the emotional feature set Q. The modal features extract the cyclical and trend characteristics of the production demand of new energy vehicles through EEMD decomposition, while the emotional features reflect the immediate reaction of the market and consumer sentiment. The multiplication dot product cross-link operation combines these two types of features and can capture the interaction effect between the two in demand forecasting. Specifically, positive emotions may accelerate the growth of demand, while negative emotions may lead to a decrease in demand.

[0063] (4) Total word vector set Multiplication and dot product crossover with the sentiment feature set Q can more accurately capture the consumer demand trend for new energy vehicles. Specifically, positive consumer reviews and news may indicate an increase in demand for new energy vehicles, and vice versa.

[0064] In step S4, the second fusion features are used as inputs of the pre-trained N-BEATS network and the TFT network respectively to obtain corresponding output results.

[0065] This step is based on the basic framework of the N-BEATS network and the TFT network. After pre-training, it is directly used in this step to generate parallel output results. That is, the embodiment of the present invention mainly emphasizes that the model input is the second fusion feature obtained in the above step. , but no relevant improvements were made to the structures of the above two network models, so the specific generation process of the output results will not be introduced here.

[0066] N-BEATS (Neural Basis Expansion Analysis for Interpretable Time Series Forecasting) is a neural network model specifically designed for time series forecasting. It enables efficient and accurate forecasting of time series, and provides interpretable information. A key advantage of N-BEATS is its adaptive selection of basis functions at different levels and time windows, enabling the model to capture various fluctuations and periodicities in time series.

[0067] The TFT (Temporal Fusion Transformer) model is an advanced deep learning model designed specifically for time series forecasting. It combines multiple neural network mechanisms to process complex relationships in time series data and uses canonical components to construct feature representations for each input type (i.e., static, known input, observed input), enabling it to achieve high predictive performance on a wide range of problems.

[0068] In step S5, the output results of the two networks are fused, and the sliding average method is used to optimize the fusion result to obtain the final new energy vehicle production demand forecast result.

[0069] To combine the outputs of the N-BEATS network and the TFT network, this step uses a weighted fusion method. To ensure the stability and accuracy of the prediction results, this step also uses a sliding average method to optimize the fusion results to obtain the final new energy vehicle production demand forecast result F.

[0070] In step S6, the final new energy vehicle production demand forecast result is used as the input of the large language model to obtain personalized production recommendations.

[0071] As a further improvement, this step introduces a large language model such as ChatGPT to obtain personalized production recommendations that can be used as a reference for new energy vehicle manufacturers.

[0072] For example, in this step, based on the final new energy vehicle production demand forecast result F, production demand forecasts at different time scales and production demand for different vehicle configurations can be analyzed. The production demand results for different vehicle configurations at different time scales are then input into ChatGPT. Based on the manufacturer's specific circumstances, personalized production recommendations are then generated from the perspectives of trend, supply chain, capacity balance, and resource optimization. Specifically, the following are analyzed: first, production demand growth / decline trends are analyzed to determine whether production demand fluctuates during different seasons or holidays. For example, whether demand increases or decreases during specific holidays or quarters, which has a significant impact on production scheduling and inventory management; second, potential supply chain bottlenecks (such as battery supply and parts procurement) are analyzed, and whether current production lines can meet future demand growth is assessed. The resulting analysis predicts the need for expansion or upgrades to production facilities; third, production capacity is analyzed to balance production capacity across different production lines, particularly during new model releases or hot sales, to ensure that resources are properly allocated to optimize overall production efficiency.

[0073] At this point, the embodiment of the present invention has completed the entire process of the new energy vehicle production demand forecasting method based on multimodal fusion.

[0074] Example 2: The embodiment of the present invention provides a new energy vehicle production demand forecasting system based on multimodal fusion, comprising: A data acquisition module for acquiring multi-source multi-modal data; the multi-source multi-modal data includes a set of numerical indicators of new energy vehicles and a text-based dataset, wherein the text-based dataset includes a set of news texts, social media comments, and industry reports; A data preprocessing module is configured to obtain a numerical feature set and a modal feature set based on the numerical indicator set; obtain a total word vector set based on the news text set, social media comments, and industry reports; and obtain a sentiment feature set based on the news text set; A feature fusion module is configured to concatenate the numerical features, the modal feature set, the total word vector set, and the sentiment feature set to obtain a first fused feature; and to process the first fused feature using a multiplication cross method to obtain a second fused feature; A result prediction module is used to use the second fusion feature as the input of the pre-trained N-BEATS network and the TFT network respectively to obtain the corresponding output results; The result optimization module is used to fuse the output results of the two networks and optimize the fusion results using the sliding average method to obtain the final new energy vehicle production demand forecast results.

[0075] In an optional implementation manner, the new energy vehicle production demand forecasting system provided by the embodiment of the present invention further includes: The recommendation acquisition module is used to use the final new energy vehicle production demand forecast results as input to the large language model to obtain personalized production recommendations.

[0076] Example 3: An embodiment of the present invention provides a storage medium storing a computer program for predicting the production demand of new energy vehicles based on multimodal fusion, wherein the computer program enables a computer to execute the method for predicting the production demand of new energy vehicles as described in Example 1.

[0077] Example 4: An embodiment of the present invention provides an electronic device, including: One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the new energy vehicle production demand forecasting method as described in Example 1.

[0078] It can be understood that the new energy vehicle production demand forecasting system based on multimodal fusion, storage medium and electronic device provided in the embodiments of the present invention correspond to the new energy vehicle production demand forecasting method based on multimodal fusion provided in the embodiments of the present invention. The explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the method and will not be repeated here.

[0079] In summary, compared with the existing technology, the present invention has the following beneficial effects: 1. In an embodiment of the present invention, multi-source multi-modal data is first acquired to comprehensively reflect the influencing factors of new energy vehicle production demand; secondly, at the feature mining level, based on the acquired multi-source multi-modal data, original fusion features including numerical features, modal feature sets, total word vector sets, and sentiment feature sets are constructed, and then processed using the multiplication cross method to fully mine the fusion features that are most predictive of production demand; thirdly, the pre-trained N-BEATS network and TFT network are used to obtain the corresponding output results to improve the interpretability of the prediction; finally, the fusion and sliding average method are used to further process the output results of the two networks to ensure the accuracy and stability of the new energy vehicle production demand prediction.

[0080] 2. The embodiments of the present invention help new energy vehicle manufacturers to scientifically and rationally arrange production plans, optimize inventory management, effectively reduce production costs and improve production efficiency, thereby enhancing the manufacturers' competitiveness in the market.

[0081] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0082] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A new energy vehicle production demand forecasting method based on multimodal fusion, characterized in that: include: Acquire multi-source multi-modal data; the multi-source multi-modal data includes a set of numerical indicators and a textual data set of new energy vehicles, wherein the textual data set includes a news text set, social media comments, and industry reports; Based on the numerical indicator set, respectively obtaining a numerical feature set and a modal feature set; based on the news text set, social media comments and industry reports, obtaining a total word vector set; and based on the news text set, obtaining a sentiment feature set; Concatenating the numerical features, the modal feature set, the total word vector set, and the sentiment feature set to obtain a first fusion feature; and processing the first fused features using a multiplication cross method to obtain a second fused feature; The second fusion feature is used as the input of the pre-trained N-BEATS network and the TFT network respectively to obtain the corresponding output results; The output results of the two networks are fused, and the sliding average method is used to optimize the fusion results to obtain the final new energy vehicle production demand forecast results.

2. The method for predicting production demand of new energy vehicles according to claim 1, characterized in that: include: The final new energy vehicle production demand forecast results are used as input to the large language model to obtain personalized production recommendations.

3. The method for predicting production demand of new energy vehicles according to claim 1, characterized in that: Based on the numerical indicator set, a numerical feature is obtained, including: The missing values ​​in the numerical indicator set are processed by the mean filling method, and the abnormal values ​​in the numerical indicator set are processed by the truncation or replacement method of the box plot analysis to obtain complete numerical indicator data; Performing Min-Max normalization processing on the complete numerical indicator data to obtain the numerical features; And / or based on the numerical indicator set, obtaining a modal feature set, including: Performing a Spearman correlation analysis on the current indicator and other indicators in the numerical indicator set to determine the indicator with the highest correlation; Add several groups of different white noises to the indicator with the highest correlation to obtain a set of noisy signals; Perform empirical mode decomposition on each noisy signal to obtain the corresponding intrinsic mode function components; The corresponding intrinsic mode function components are averaged respectively to obtain the collective empirical mode decomposition results as the corresponding modal features; All indicators in the numerical indicator set are traversed to obtain the modal feature set.

4. The method for predicting production demand of new energy vehicles according to claim 1, wherein: The method of obtaining a total word vector set based on the news text collection, social media comments, and industry reports includes: Perform text cleaning and standardization on the news text collection, social media comments, and industry reports; We extract a preset number of words from the standardized news text collection, social media comments, and industry reports, and use the FinBERT model to obtain the corresponding word embeddings to obtain a set of news word vectors, a set of social media comment word vectors, and a set of industry report word vectors. The news word vector set, social media comment word vector set, and industry report word vector set are merged, and the principal component analysis method is used to reduce the dimension of the merging results to obtain the total word vector set.

5. The method for predicting production demand of new energy vehicles according to claim 1, wherein: The step of obtaining a sentiment feature set based on the news text set includes: Performing text cleaning and standardization on the news text collection; The SnowNLP model is used to convert each standardized news text set into a sentiment score. After traversing all news text sets, the sentiment feature set is obtained.

6. The method for predicting production demand of new energy vehicles according to claim 1, wherein: The adopting the multiplication cross method to process the first fused feature to obtain the second fused feature includes: For the numerical features and modal feature sets, numerical features and total word vector sets, numerical features and sentiment feature sets, modal feature sets and sentiment feature sets, and total word vector sets and sentiment feature sets in the first fusion features, the multiplication cross method is used to obtain the corresponding interaction features, and all the interaction features are spliced ​​to obtain the second fusion features.

7. A new energy vehicle production demand forecasting system based on multimodal fusion, characterized in that: include: A data acquisition module for acquiring multi-source multi-modal data; the multi-source multi-modal data includes a set of numerical indicators of new energy vehicles and a text-based dataset, wherein the text-based dataset includes a set of news texts, social media comments, and industry reports; A data preprocessing module, configured to obtain a numerical feature set and a modal feature set based on the numerical indicator set; Based on the news text collection, social media comments, and industry reports, obtain a total word vector set; and based on the news text collection, obtain a sentiment feature set; A feature fusion module, configured to combine the numerical features, the modal feature set, the total word vector set, and the sentiment feature set to obtain a first fused feature; and processing the first fused features using a multiplication cross method to obtain a second fused feature; A result prediction module is used to use the second fusion feature as the input of the pre-trained N-BEATS network and the TFT network respectively to obtain the corresponding output results; The result optimization module is used to fuse the output results of the two networks and optimize the fusion results using the sliding average method to obtain the final new energy vehicle production demand forecast results.

8. The new energy vehicle production demand forecasting system according to claim 7, characterized in that: include: The recommendation acquisition module is used to use the final new energy vehicle production demand forecast results as input to the large language model to obtain personalized production recommendations.

9. A storage medium, characterized in that: It stores a computer program for predicting the production demand of new energy vehicles based on multimodal fusion, wherein the computer program enables the computer to execute the new energy vehicle production demand prediction method according to any one of claims 1 to 6.

10. An electronic device, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the new energy vehicle production demand forecasting method according to any one of claims 1 to 6.

Images