Electric vehicle scale prediction method based on AI large model and multi-source information fusion

Abstract

The invention relates to the technical field of scale prediction, in particular to an electric vehicle scale prediction method based on an AI large model and multi-source information fusion, and the method comprises the following steps: collecting related data affecting the scale of an electric vehicle from a plurality of heterogeneous data sources; fusing the features from different data sources by using a multi-modal fusion technology; based on the fused features, constructing an AI large model with an uncertainty quantification capability; and carrying out hierarchical prediction based on the AI large model, and respectively predicting the scales of the electric vehicle in different areas, different vehicle types and different charging infrastructure scenes. According to the method, uncertainty measurement of the prediction result is provided by analyzing the model parameters and the output probability distribution, so that the prediction result is more credible and explainable, comprehensive evaluation under different scenes is provided, and powerful support is provided for decisions in the aspects of regional planning, vehicle model production, charging infrastructure construction and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of scale prediction, and in particular to an electric vehicle scale prediction method based on an AI large model and multi-source information fusion. Background Art

[0002] With the global emphasis on environmental protection and sustainable development, the electric vehicle (EVs) market has experienced significant growth in the past few years. EVs not only help reduce carbon emissions, but also promote the rapid development of new energy technologies. However, as the market expands, how to accurately predict the market size of electric vehicles has become an important basis for government policy making, corporate market layout planning and infrastructure construction.

[0003] Traditional electric vehicle scale forecasting methods usually rely on a single data source or simple statistical models, such as linear regression or time series analysis. These methods have significant limitations when dealing with complex, multi-source data and cannot effectively deal with the multiple factors that affect the scale of electric vehicles, including economic policies, market demand, technological progress, charging infrastructure construction, consumer preferences, and environmental regulations. These complex factors have significant heterogeneity and nonlinear relationships. Traditional methods cannot fully capture and quantify these relationships, resulting in insufficient accuracy and robustness of the forecast results.

[0004] With the rapid development of artificial intelligence technology, especially deep learning and large-scale language models, AI models based on multi-source information fusion provide a new way to solve complex prediction problems. However, existing AI models still have many shortcomings in dealing with the fusion of multimodal data, uncertainty quantification, and model interpretability. In particular, the effective fusion of heterogeneous data sources, accurate predictions at different levels, and simulation analysis under different policy and market scenarios have not been fully resolved under the current technical framework. Summary of the invention

[0005] Based on the above objectives, the present invention provides an electric vehicle scale prediction method based on AI large model and multi-source information fusion.

[0006] The electric vehicle scale prediction method based on AI big model and multi-source information fusion includes the following steps:

[0007] S1, multimodal data collection: collect relevant data affecting the scale of electric vehicles from multiple heterogeneous data sources, including economic policies, market demand, technological progress, charging infrastructure construction, consumer preferences and environmental regulations;

[0008] S2, multimodal feature fusion and selection: Multimodal fusion technology is used to fuse features from different data sources. The fusion process uses a heterogeneous attention mechanism to identify and highlight key features in different modal data;

[0009] S3, model construction and uncertainty quantification: Based on the fused features, a large AI model with uncertainty quantification capability is constructed. The model includes a Bayesian neural network (BNN) and a variational autoencoder (VAE) to capture potential uncertainties in the prediction process. By analyzing the probability distribution of the model's internal parameters and outputs, uncertainty measurement of the prediction results is provided.

[0010] S4, hierarchical prediction: hierarchical prediction based on AI big model, predicting the scale of electric vehicles in different regions, different models and different charging infrastructure scenarios.

[0011] Optionally, the multimodal feature fusion and selection in S2 specifically includes:

[0012] S21, Feature extraction and encoding: Corresponding feature extraction methods are used to process the features of different data sources, including using random forests to extract features from structured data, using embedded language models to generate text vectors for text data, using convolutional neural networks (CNNs) to extract visual features from image data, and using long short-term memory networks (LSTMs) to extract temporal features from time series data;

[0013] S22, multimodal feature alignment: align the features extracted from different modal data sources, including time synchronization and spatial registration, to ensure data consistency in the subsequent fusion process; the goal of multimodal feature alignment is to ensure that the features from different modalities are consistent in the same time, spatial dimension and scale. The modal features after feature extraction are F1, F2, ..., F k , the aligned feature is recorded as F i ′.

[0014] S23, application of heterogeneous attention mechanism: using the heterogeneous attention mechanism, different weights are assigned according to the importance of each modal data, and key features that have a key impact on the electric vehicle scale prediction are identified and highlighted. The heterogeneous attention mechanism includes an attention model based on weighted summation, which calculates attention scores based on different modal data to improve the fusion effect of multi-source information;

[0015] S24, feature fusion: The features weighted by the attention mechanism are fused, and a fusion strategy is implemented based on the cross-modal transformer to embed the features of different modalities into a unified high-dimensional space to generate the final fused feature representation.

[0016] Optionally, a heterogeneous attention mechanism is used to assign the importance of different modal features. For each modal feature F i ′, by calculating the attention weight α i To identify key features:

[0017] Weighted sum attention: For each modality i, calculate its attention weight α i :

[0018] Where W i is the weight matrix;

[0019] Initial fusion feature generation: Initial fusion feature F fusion By weighted summation we get:

[0020]

[0021] Optionally, the cross-modal transformer embeds the features of different modalities into a unified high-dimensional space and performs final feature fusion to convert the preliminary fused features F fusion As input features, the output of the cross-modal transformer is F out ;

[0022] Based on the multi-head self-attention mechanism, calculate the attention weight of each head:

[0023] Where Q i , K i , V i are query, key, and value, respectively. is the corresponding weight matrix;

[0024] Compute the attention output:

[0025] Finally, the multi-head attention output is fused:

[0026] F out =Concat(Attention1,...,Attention h ), where h is the number of heads, d k The dimension of the key.

[0027] Optionally, the construction of the Bayesian neural network BNN is based on the final fusion feature F out In BNN, model parameters are not fixed values, but are expressed as probability distribution p(θ). Through maximum likelihood estimation and variational inference, the posterior distribution p(θ|D) is calculated, where D is the training data. During model training, the probability distribution of parameters is updated by multiple sampling of training data to capture the uncertainty in the parameters: in, is an approximate variational distribution, with variational parameters control.

[0028] Optionally, the variational autoencoder VAE is used to capture the potential uncertainty in the feature space. The VAE consists of an encoder and a decoder. The encoder takes the input feature F out The probability distribution q(z|F out ), the decoder samples from this probability distribution to generate reconstructed features By minimizing the reconstruction loss and KL divergence, the VAE is trained to capture the underlying uncertainty in the features.

[0029] Optionally, the uncertainty measure is obtained by performing uncertainty quantification analysis on the outputs of BNN and VAE. For BNN, the predicted distribution of the model output is calculated. Its variance represents the uncertainty of the prediction. For VAE, the distribution parameters of the latent variable z (mean μ and variance σ) are calculated. 2 ), quantifying the uncertainty in the feature space:

[0030]

[0031] Optionally, S3 also includes comprehensive uncertainty analysis, combining the outputs of BNN and VAE, and comprehensively analyzing the two uncertainties by weighted summation to obtain the final uncertainty

[0032] Uncertainty final :

[0033] Uncertainty final =α·Uncertainty BNN +β·Uncertainty VAE , where α and β are adjustment weights, which are adjusted according to model performance and actual needs.

[0034] Optionally, the hierarchical prediction in S4 specifically includes:

[0035] For prediction tasks at different levels, feature subsets F region, F vehicle type, and F infrastructure at different levels are extracted from the final fused feature Fout, representing different regional levels, different vehicle type levels, and different charging infrastructure scenario levels, respectively:

[0036] Hierarchical prediction model construction: Apply the trained AI big model to prediction tasks at different levels, including electric vehicle scale prediction in different regions, different models and different charging infrastructure scenarios, and generate prediction outputs for different levels, including different regional outputs Output of different models

[0037] Output of different charging infrastructure scenarios

[0038]

[0039] Among them, AI Model represents the AI ​​big model.

[0040] Optionally, the AI ​​Model includes a Bayesian neural network BNN and a variational autoencoder VAE. For each feature subset at each level, prediction is performed based on the BNN. The BNN accepts the feature subsets F region, F vehicle type, and Finfrastructure as input and generates prediction outputs for each level. and accompanied by uncertainty quantification;

[0041] The latent space of each feature subset is modeled using a variational autoencoder (VAE). The output of the VAE includes the reconstructed features. and the distribution parameters of the latent variables μregion, σregion, μvehicle type, σvehicle type,

[0042] μinfrastructure,σinfrastructure;

[0043] Quantify the uncertainty at each level by calculating the variance of the latent variables;

[0044] Combining the outputs of BNN and VAE, the prediction results of each level are comprehensively analyzed. Through weighted calculation, the uncertainties of BNN and VAE are combined to obtain the final hierarchical prediction output and uncertainty measurement.

[0045] Beneficial effects of the present invention:

[0046] The present invention utilizes heterogeneous attention mechanism and cross-modal transformer technology to achieve effective fusion of multiple heterogeneous data sources and generate a unified high-dimensional feature representation. This fusion method can highlight and capture key features in different data modalities, ensure the robustness of the prediction model when processing complex data, and further introduce an uncertainty quantification mechanism by combining Bayesian neural network (BNN) and variational autoencoder (VAE). By analyzing the probability distribution of model parameters and output, an uncertainty measure of the prediction result is provided, making the prediction result more credible and explainable.

[0047] The present invention selects feature subsets and makes independent predictions for different levels (region, vehicle type, charging infrastructure) in hierarchical prediction, thereby ensuring the accuracy of the prediction results. By introducing a scenario simulator, the scale development trend of electric vehicles under different policies and market scenarios is simulated. The method combining hierarchical prediction with scenario simulation provides a comprehensive evaluation under different scenarios, thus providing strong support for decision-making in regional planning, vehicle type production, and charging infrastructure construction.

[0048] The present invention combines multimodal fusion, Bayesian neural network (BNN) and variational autoencoder (VAE) in the AI ​​large model to achieve accurate prediction of complex and multi-source data, which not only improves the accuracy of electric vehicle scale prediction, but also enhances the adaptability of the model in the face of different data patterns and environmental changes. By comprehensively considering the prediction results and uncertainty analysis at all levels, it provides decision makers with more comprehensive decision support, significantly improving the practical application value of electric vehicle market prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 A schematic diagram of a prediction method flow chart of an embodiment of the present invention;

[0051] Figure 2 Schematic diagram of multimodal feature fusion and selection steps according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0053] It should be noted that the references to "one embodiment", "an embodiment", "an exemplary embodiment", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0054] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0055] like Figure 1-Figure 2 As shown, the electric vehicle scale prediction method based on AI big model and multi-source information fusion includes the following steps:

[0056] S1, multimodal data collection: collect relevant data affecting the scale of electric vehicles from multiple heterogeneous data sources, including economic policies, market demand, technological progress, charging infrastructure construction, consumer preferences and environmental regulations;

[0057] a) Economic policy data collection: obtain the latest economic policy information through public data platforms and government statistical databases;

[0058] b) Market demand data collection: Collect market demand data from market research reports, sales data, consumer surveys and other sources, and normalize time series data to eliminate the impact of fluctuations and ensure data continuity and consistency;

[0059] c) Data collection on technological progress: Obtain information on technological progress related to electric vehicles through patent databases, academic literature, and industry reports, and perform structured processing on text data;

[0060] d) Charging infrastructure construction data collection: Obtaining infrastructure construction data such as charging station distribution and number of charging piles through government planning documents and industry databases;

[0061] e) Consumer preference data collection and preprocessing: Obtain consumer preference data from social media, online surveys and user reviews;

[0062] f) Environmental regulation data collection and preprocessing: Obtain environmental regulation information related to electric vehicles from environmental policy documents and regulatory databases, and use rule extraction and ontology modeling technology to structure the regulatory clauses.

[0063] S2, multimodal feature fusion and selection: Multimodal fusion technology is used to fuse features from different data sources. The fusion process uses a heterogeneous attention mechanism to identify and highlight key features in different modal data;

[0064] S3, model construction and uncertainty quantification: Based on the fused features, a large AI model with uncertainty quantification capability is constructed. The model includes a Bayesian neural network (BNN) and a variational autoencoder (VAE) to capture potential uncertainties in the prediction process. By analyzing the probability distribution of the model's internal parameters and outputs, it provides uncertainty measures for the prediction results. Bayesian neural networks (BNN) and variational autoencoders (VAE) are both complex models in deep learning. BNN models parameter uncertainty through probability distribution, and VAE captures uncertainty in feature space through encoding and decoding. Combining BNN and VAE to conduct a comprehensive analysis of the uncertainty of the prediction results enables the model to not only make accurate predictions, but also quantify and explain the uncertainty in the predictions. This multi-model fusion method can handle complex multi-source information and adapt to the fusion and inference of multimodal data features.

[0065] S4, hierarchical prediction: hierarchical prediction based on AI big model, predicting the scale of electric vehicles in different regions, different models and different charging infrastructure scenarios.

[0066] The multimodal feature fusion and selection in S2 specifically include:

[0067] S21, Feature extraction and encoding: Corresponding feature extraction methods are used to process the features of different data sources, including using random forests to extract features from structured data, using embedded language models to generate text vectors for text data, using convolutional neural networks (CNN) to extract visual features from image data, and using long short-term memory networks (LSTM) to extract temporal features from time series data; the details are as follows:

[0068] 1. Feature extraction of structured data: For structured data (market demand, charging infrastructure construction), the random forest algorithm is used to extract features. Assume that the data set is X∈R n×m , where n is the number of samples and m is the number of features, the output of the random forest model is the feature importance:

[0069] Where T is the number of decision trees, I t is the feature importance score of the t-th decision tree.

[0070] 2. Embedded language model encoding of text data: For text data (economic policies and consumer preferences), the BERT pre-trained model is used to convert text data into embedded vectors. Assume that the input text sequence is X text ={x1, x2, ..., x n}, the output embedding vector of the BERT model is E text ∈R n×d , where d is the dimension of the embedding vector: Etext =BERT(X text ).

[0071] 3. Convolutional neural network feature extraction of image data: For image data (charging facility distribution map), convolutional neural network (CNN) is used to extract features. Assume that the input image is I∈R H×W×C , where H, W are the height and width of the image, C is the number of channels, and the output feature map of CNN is F∈R h×w×c : F = CNN(I), where h, w and c are the height, brightness and number of channels of the output feature map respectively.

[0072] 4. Feature extraction of time series data: For time series data (market demand time series), use long short-term memory network (LSTM) to extract features. Assume that the input time series data is X time ={x1, x2, ..., x T}, the output of LSTM is the hidden state H t =LSTM(X time ), where t is the time step, d h is the dimension of the hidden state.

[0073] S22, multimodal feature alignment: align the features extracted from different modal data sources, including time synchronization and spatial registration, to ensure data consistency in the subsequent fusion process; the goal of multimodal feature alignment is to ensure that the features from different modalities are consistent in the same time, spatial dimension and scale. The modal features after feature extraction are F1, F2, ..., F k , the aligned feature is recorded as F i ',but:

[0074] Time synchronization: For time series data, interpolate or sample the time steps so that different data modalities are aligned to the same time step t: F i ′(t)=Interpolate(F i ), where Interpolate represents an interpolation operation;

[0075] Spatial registration: For image data, the features of different modalities are aligned to the same spatial coordinate system through spatial transformation: F i ′(t)=Spatial Transform(F i ), where Spatial Transform represents spatial transformation.

[0076] S23, Application of heterogeneous attention mechanism: Using the heterogeneous attention mechanism, different weights are assigned according to the importance of each modal data to identify and highlight the key features that have a key impact on the electric vehicle scale prediction. The heterogeneous attention mechanism includes an attention model based on weighted summation, which calculates the attention score based on different modal data to improve the fusion effect of multi-source information;

[0077] S24, feature fusion: The features weighted by the attention mechanism are fused, and a fusion strategy is implemented based on the cross-modal transformer to embed the features of different modalities into a unified high-dimensional space to generate the final fused feature representation.

[0078] The main purpose of the heterogeneous attention mechanism is to identify and highlight the key features in different modal data in the initial fusion stage. By weighted summation, it focuses on weighted combination of features from multiple sources according to their importance to generate a preliminary fusion feature. The heterogeneous attention mechanism is used to allocate the importance of different modal features. For each modal feature F i ′, by calculating the attention weight α i To identify key features:

[0079] Weighted sum attention: For each modality i, calculate its attention weight α i :

[0080] Where W i is the weight matrix;

[0081] Initial fusion feature generation: Initial fusion feature F fusion By weighted summation we get:

[0082]

[0083] Based on the preliminary fusion features, the complex attention mechanism and feature interaction are used to further improve the expression ability of the fusion features and capture more complex relationships between modalities. The cross-modal transformer embeds the features of different modalities into a unified high-dimensional space and performs the final feature fusion to convert the preliminary fusion features F fusion As input features, the output of the cross-modal transformer is F out ;

[0084] Based on the multi-head self-attention mechanism, calculate the attention weight of each head:

[0085] Where Q i , K i , V i are query, key, and value, respectively. is the corresponding weight matrix;

[0086] Compute the attention output:

[0087] Finally, the multi-head attention output is fused:

[0088] F out =Concat(Attention1,...,Attention h ), where h is the number of heads, d k The dimension of the key.

[0089] The construction of Bayesian neural network BNN is based on the final fusion feature F out In BNN, model parameters are not fixed values, but are expressed as probability distribution p(θ). Through maximum likelihood estimation and variational inference, the posterior distribution p(θ|D) is calculated, where D is the training data. During model training, the probability distribution of parameters is updated by multiple sampling of training data to capture the uncertainty in the parameters:

[0090] in, is an approximate variational distribution, with variational parameters

[0091] control.

[0092] Specifically, the detailed construction and training of the Bayesian neural network are as follows:

[0093] 1. Input features: the final fused features F out is a feature vector with high-dimensional representation generated by the cross-modal transformer. Assume that F out The dimension of is d, that is

[0094] 2. BNN model construction: Bayesian neural network is a deep neural network, in which weight W and bias b are modeled as probability distributions rather than fixed values. The structure of BNN is a fully connected network with L layers, and the input of each layer is h (l-1) , the output is h (l) , and the network parameters (weights and biases) are W (l) and b (l) :h (l) =σ(W (l) ·h (l-1) +b (l) ),in:

[0095] h (0) =F out It is the input layer, representing the final fused features.

[0096] W (l) ~N(μ (l), σ (l)2 ) is the weight matrix of the Ith layer, which obeys the normal distribution.

[0097] is the bias vector of the Ith layer, which also follows a normal distribution.

[0098] σ(·) is the ReLU activation function.

[0099] 3. Posterior distribution calculation: During the training process, using the training data:

[0100] The posterior distribution of the network parameters is approximated, with the goal of maximizing the following likelihood function (i.e., achieved through variational inference):

[0101] in:

[0102] θ: The set of all parameters of the BNN model, including weights and biases.

[0103] It is the parameter set of all network layers.

[0104] p(D|θ) is the likelihood function, which represents the probability of data given the parameters.

[0105] p(θ) is the prior distribution of the parameters.

[0106] Through variational inference, the posterior distribution p(θ|D) is approximated as a variational distribution And by optimizing the variational parameters to approximately maximize the log-likelihood.

[0107] 4. Prediction output: BNN output prediction It is also a probability distribution, whose expected value and variance are calculated by multiple sampling to capture the prediction uncertainty:

[0108] Where M is the number of sampling times, is the output prediction distribution of the mth sampling.

[0109] Through the above steps, the BNN model can not only make predictions based on the final fused feature Fout, but also quantify the uncertainty of the prediction, which provides a solid foundation for subsequent hierarchical predictions.

[0110] Variational Autoencoder VAE is used to capture the potential uncertainty in the feature space. VAE consists of an encoder and a decoder. The encoder takes the input feature F out Probability distribution mapped to latent variables

[0111] q(z|F out), the decoder samples from this probability distribution to generate reconstructed features By minimizing the reconstruction loss and KL divergence, VAE is trained to capture the underlying uncertainty in the features, expressed as:

[0112] Here, p(z) is the prior distribution, usually the standard normal distribution.

[0113] Uncertainty measures are performed by quantifying the uncertainty of the outputs of BNN and VAE. For BNN, the predicted distribution of the model output is calculated. Its variance represents the uncertainty of the prediction. For VAE, the distribution parameters of the latent variable z (mean μ and variance σ) are calculated. 2 ), quantifying the uncertainty in the feature space:

[0114]

[0115] S3 also includes comprehensive uncertainty analysis, combining the outputs of BNN and VAE, and comprehensively analyzing the two uncertainties through weighted summation to obtain the final uncertainty Uncertainty final :

[0116] Uncertainty final =α·Uncertainty BNN +β·Uncertainty VAE , where α and β are adjustment weights, which are adjusted according to model performance and actual needs.

[0117] The hierarchical prediction in S4 specifically includes:

[0118] For prediction tasks at different levels, feature subsets F region, F vehicle type, and F infrastructure at different levels are extracted from the final fused feature Fout, representing different regional levels, different vehicle type levels, and different charging infrastructure scenario levels, respectively:

[0119] F region=Select Features(Fout, Region Data); Region Data is

[0120] Regional data;

[0121] F veh icletype=Select Features(Fout, Vehicle Type Data);

[0122] Vehicle Type Data is vehicle type data;

[0123] F infrastructure=Select Features(Fout, Infrastructure Data);

[0124] Infrastructure Data is charging infrastructure data;

[0125] Among them, Select Features represents the feature selection operation of selecting a subset of features related to a specific level (region, vehicle model, charging infrastructure) from the final fused features;

[0126] Hierarchical prediction model construction: Apply the trained AI big model to prediction tasks at different levels, including electric vehicle scale prediction in different regions, different models and different charging infrastructure scenarios, and generate prediction outputs for different levels, including different regional outputs Output of different models

[0127] Output of different charging infrastructure scenarios

[0128]

[0129] Among them, AI Model represents the AI ​​big model.

[0130] The AI ​​Model includes a Bayesian neural network (BNN) and a variational autoencoder (VAE). For each feature subset at each level, predictions are made based on the BNN. The BNN accepts the feature subsets F region, F vehicle type, and F infrastructure as inputs and generates prediction outputs for each level. With uncertainty quantification, the prediction output is calculated as follows:

[0131]

[0132] At the same time, the uncertainty is quantified by calculating the variance of the output distribution:

[0133]

[0134]

[0135] The latent space of each feature subset is modeled using a variational autoencoder (VAE). The output of the VAE includes the reconstructed features. And the distribution parameters of the latent variables μregion, σregion, μveh icle type, σveh icle type,

[0136] μinfrastructure, σinfrastructure:

[0137]

[0138] Quantify the uncertainty at each level by calculating the variance of the latent variables;

[0139] Combining the outputs of BNN and VAE, the prediction results of each layer are comprehensively analyzed. Through weighted calculation, the uncertainty of BNN and VAE is combined to obtain the final hierarchical prediction output and uncertainty measurement:

[0140]

[0141] Uncertaintyf inal,region=αBNN·UncertaintyBNN,region+βVAE·σregion 2 ;

[0143] Among them, αBNN and βVAE are adjustment weights, which are adjusted according to model performance and actual needs. Similarly, the other levels (vehicle types and charging infrastructure) are calculated in the same way.

[0144] A scenario simulator is also introduced to simulate the scale development trend of electric vehicles under different policies and market scenarios. A scenario simulator is built to simulate the scale development trend of electric vehicles under different policies and market scenarios. The scenario simulator is based on a variety of scenarios that may occur in reality (such as policy changes, market demand fluctuations, technological advances, etc.) and generates feature inputs under multiple scenarios. And input it into the AI ​​big model to generate prediction results for each scenario

[0145] Among them, s represents different scenario simulations.

[0146] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0147] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. The electric vehicle scale prediction method based on AI large model and multi-source information fusion is characterized by: The following steps are involved: S1, multimodal data collection: collect relevant data affecting the scale of electric vehicles from multiple heterogeneous data sources, including economic policies, market demand, technological progress, charging infrastructure construction, consumer preferences and environmental regulations; S2, multimodal feature fusion and selection: Multimodal fusion technology is used to fuse features from different data sources. The fusion process uses a heterogeneous attention mechanism to identify and highlight key features in different modal data; S3, model construction and uncertainty quantification: Based on the fused features, a large AI model with uncertainty quantification capability is constructed. The model includes a Bayesian neural network (BNN) and a variational autoencoder (VAE) to capture potential uncertainties in the prediction process. By analyzing the probability distribution of the model's internal parameters and outputs, uncertainty measurement of the prediction results is provided. S4, hierarchical prediction: hierarchical prediction based on AI big model, predicting the scale of electric vehicles in different regions, different models and different charging infrastructure scenarios.

2. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 1 is characterized in that: The multimodal feature fusion and selection in S2 specifically include: S21, Feature extraction and encoding: Corresponding feature extraction methods are used to process the features of different data sources, including using random forests to extract features from structured data, using embedded language models to generate text vectors for text data, using convolutional neural networks to extract visual features from image data, and using long short-term memory networks to extract temporal features from time series data; S22, multimodal feature alignment: align the features extracted from different modal data sources, including through time synchronization and spatial registration. The goal of multimodal feature alignment is to ensure that the features from different modalities are consistent in the same time, spatial dimension and scale. The modal features after feature extraction are F1, F2, ..., F k , the aligned feature is recorded as F i ′; S23, application of heterogeneous attention mechanism: using the heterogeneous attention mechanism, different weights are assigned according to the importance of each modal data, and key features that have a key impact on the electric vehicle scale prediction are identified and highlighted. The heterogeneous attention mechanism includes an attention model based on weighted summation, which calculates attention scores based on different modal data to improve the fusion effect of multi-source information; S24, feature fusion: The features weighted by the attention mechanism are fused, and a fusion strategy is implemented based on the cross-modal transformer to embed the features of different modalities into a unified high-dimensional space to generate the final fused feature representation.

3. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 2 is characterized in that: The heterogeneous attention mechanism is used to allocate the importance of different modal features. For each modal feature F i ′, by calculating the attention weight α i To identify key features: Weighted sum attention: For each modality i, calculate its attention weight α i : Where W i is the weight matrix; Initial fusion feature generation: Initial fusion feature F fusion By weighted summation we get:

4. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 3 is characterized in that: The cross-modal transformer embeds the features of different modalities into a unified high-dimensional space and performs final feature fusion to convert the preliminary fusion feature F fusion As input features, the output of the cross-modal transformer is F out ; Based on the multi-head self-attention mechanism, calculate the attention weight of each head: Where Q i , K i , V i are query, key, and value, respectively. is the corresponding weight matrix; Compute the attention output: Finally, the multi-head attention output is fused: F out =Concat(Attention1,...,Attention h ), where h is the number of heads, d k The dimension of the key.

5. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 4 is characterized in that: The Bayesian neural network BNN is constructed based on the final fusion feature F out In BNN, model parameters are represented as probability distribution p(θ). Through maximum likelihood estimation and variational inference, the posterior distribution p(θ|D) is calculated, where D is the training data. During model training, the probability distribution of parameters is updated by multiple sampling of training data to capture the uncertainty in the parameters: in, is an approximate variational distribution, with variational parameters control.

6. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 5 is characterized in that: The variational autoencoder VAE is used to capture the potential uncertainty in the feature space. VAE consists of an encoder and a decoder. The encoder takes the input feature F out The probability distribution q(z|F out ), the decoder samples from this probability distribution to generate the reconstructed features By minimizing the reconstruction loss and KL divergence, the VAE is trained to capture the underlying uncertainty in the features.

7. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 6 is characterized in that: The uncertainty measure is obtained by performing uncertainty quantification analysis on the output of BNN and VAE. For BNN, the predicted distribution of the model output is calculated. Its variance represents the uncertainty of the prediction. For VAE, the uncertainty in the feature space is quantified by calculating the variance of the distribution parameters of the latent variable z:

8. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 7 is characterized in that: S3 also includes comprehensive uncertainty analysis, combining the outputs of BNN and VAE, and comprehensively analyzing the two uncertainties by weighted summation to obtain the final uncertainty Uncertainty final : Uncertainty final =α·Uncertainty BNN +β·Uncertainty VAE , where α and β are adjustment weights.

9. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 1 is characterized in that: The hierarchical prediction in S4 specifically includes: For prediction tasks at different levels, the final fused feature F out Extract feature subsets Fregion, Fveh icle type, and Finfrastructure at different levels, which represent different regional levels, different vehicle types, and different charging infrastructure scenario levels respectively; Hierarchical prediction model construction: Apply the trained AI big model to prediction tasks at different levels, including electric vehicle scale prediction in different regions, different models and different charging infrastructure scenarios, and generate prediction outputs for different levels, including different regional outputs Output of different models Output of different charging infrastructure scenarios Among them, AI Model represents the AI ​​big model.

10. The electric vehicle scale prediction method based on AI large model and multi-source information fusion according to claim 1 is characterized in that: The AT Model includes a Bayesian neural network (BNN) and a variational autoencoder (VAE). For each feature subset at each level, predictions are made based on the BNN. The BNN accepts the feature subsets F region, F vehicle type, and Finfrastructure as inputs and generates prediction outputs for each level. and accompanied by uncertainty quantification; The latent space of each feature subset is modeled using a variational autoencoder (VAE). The output of the VAE includes the reconstructed features. and the distribution parameters of the latent variables μregion, σregion, μveh icle type, σveh icle type, μinfrastructure, σinfrastructure. The uncertainty at each level is quantified by calculating the variance of the latent variables; Combining the outputs of BNN and VAE, the prediction results of each level are comprehensively analyzed. Through weighted calculation, the uncertainties of BNN and VAE are combined to obtain the final hierarchical prediction output and uncertainty measurement.

Images