Second-hand new energy automobile price prediction method based on LSTM-GCN model

Abstract

The invention discloses a second-hand new energy automobile price prediction method based on an LSTM-GCN model. The method comprises the steps of 1, obtaining historical price time sequence data and related influence factor data of a second-hand new energy automobile market; 2, cleaning and standardizing the data; 3, analyzing a market structure based on data; 4, on the basis of the market structure, constructing a plurality of models for representing the spatial dependency relationships, and for each spatial dependency relationship, generating a plurality of corresponding adjacency matrixes by using a graph calculation method; 5, inputting historical price time sequence data, influence factor data and various adjacent matrixes into the trained price prediction model, and outputting corresponding second-hand new energy vehicle price prediction data; and 6, using second-hand new energy vehicle price prediction data output by the price prediction model as a market price prediction result. According to the method, the prediction precision is remarkably improved, and a more reliable price prediction reference can be provided for participants in the second-hand new energy automobile market.

Description

Technical Field

[0001] This invention proposes a method for predicting the price of used new energy vehicles based on the LSTM-GCN model, particularly relating to the fields of big data analysis and new energy vehicle technology. Background Technology

[0002] With increasing global environmental awareness and accelerated energy structure transformation, the used new energy vehicle market has experienced rapid growth in recent years. As a core market indicator, the accurate prediction of used new energy vehicle prices is crucial for manufacturers to formulate production strategies, consumers to make purchasing decisions, and governments to formulate policies. However, due to the unique and complex nature of the used new energy vehicle market, traditional price prediction methods face numerous challenges when applied to this field.

[0003] Current technologies for predicting the prices of used new energy vehicles still have the following shortcomings: First, most existing price prediction models are designed for traditional energy or mature markets, failing to fully consider the unique characteristics of the used new energy vehicle market, such as strong policy dependence, rapid technological iteration, and complex market structure. Second, traditional time series analysis methods, such as ARIMA and exponential smoothing, are difficult to effectively capture the nonlinear characteristics and long-term dependencies in the price fluctuations of used new energy vehicles. Third, although some studies have introduced deep learning methods such as LSTM for price prediction, these methods mainly focus on feature extraction in the time dimension, ignoring the spatial correlations between different entities in the used new energy vehicle market, such as competitive relationships, supply chain relationships, and the impact of policies and economies. Finally, existing methods face difficulties in integrating multi-source heterogeneous data, making it difficult to comprehensively consider the various factors affecting the prices of used new energy vehicles, such as vehicle type, supplier procurement amount, regional economic level, competitor vehicle prices, and market supply and demand data.

[0004] Graph Neural Networks (GNNs) are a deep learning architecture specifically designed for processing non-Euclidean spatial data. Their core idea is to aggregate topological relationships between nodes and their neighborhoods through a message-passing mechanism. GNNs excel at modeling graph structures, capturing non-linear relationships and global dependencies between nodes, and can dynamically adjust the strength of relationships between dimensions through edge weights, thus more accurately reflecting reality. Due to their powerful graph structure modeling capabilities, cross-dimensional relationship capture abilities, dynamic modeling capabilities, and ability to process heterogeneous data, by fully leveraging the graph structure characteristics of used new energy vehicle data, GNNs can provide more accurate and flexible price prediction results than traditional methods.

[0005] Therefore, there is an urgent need for a method to predict the price of used new energy vehicles that can simultaneously consider time series characteristics and spatial correlations, and effectively integrate multi-source heterogeneous data, in order to improve the accuracy and reliability of predictions and provide stronger support for relevant decision-making. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method for predicting the price of used new energy vehicles based on the LSTM-GCN model. By using the LSTM-GCN model as a foundation, this method solves the technical problem of multiple complex influencing factors affecting the price of used new energy vehicles, thereby improving the accuracy of price prediction.

[0007] The technical solution adopted in this invention is as follows: A method for predicting the price of used new energy vehicles based on the LSTM-GCN model, the steps of which are: Step S1: Obtain historical price time series data and related influencing factor data for the used new energy vehicle market; Relevant influencing factors include: vehicle positioning and configuration, supplier procurement amounts, policy and economic factors, competitor vehicle prices, power battery technology, and market supply and demand data.

[0008] Step S2: Clean and standardize the historical price time series data and related influencing factor data of the collected used new energy vehicle market; The specific method is as follows: S21. Use time series interpolation to complete missing price data, and use the isolated forest algorithm to check and correct outliers, and perform Min-Max standardization on multivariate data; S22. Aggregate daily data into monthly averages, and decompose quarterly data into monthly data through linear interpolation; and use embedding to process vehicle configuration features.

[0009] Step S3: Analyze the structure of the used new energy vehicle market based on historical price data and related influencing factors; Step S4: Based on the structure of the used new energy vehicle market, construct a variety of models to represent spatial dependencies, and generate several corresponding adjacency matrices for each spatial dependency using graph computing methods. The specific method is as follows: S41. By constructing graphs, model the four spatial dependencies of market structure: competitive relationship, supply chain relationship, policy and economic impact, and power battery technology impact, to obtain competitive relationship graph, supply chain relationship graph, policy and economic impact graph, and battery technology graph; S42. Calculate the adjacency matrix A for the competition graph, supply chain graph, policy and economic impact graph, and battery technology graph. comp A supply A policy and A battery .

[0010] The adjacency matrix A of the competition graph is calculated using the following formula. comp Elements in:

[0011] In the formula: C ij P represents the competitive intensity between brand i and brand j; i and P j P represents the average price of brand i and brand j, respectively; max Indicates the highest price in the market; S ij This indicates the degree of overlap between brand i and brand j in the target customer group.

[0012] The adjacency matrix A of the supply chain relationship graph is calculated using the following formula. supply Elements in:

[0013] In the formula: A ij Indicates the supply chain strength between brand i and supplier j; T ij T represents the amount that brand i purchases from supplier j; max This represents the maximum single purchase amount for brand i; To prevent small constants with a denominator of zero.

[0014] The adjacency matrix A of the policy-economic impact diagram is calculated using the following formula. policy Elements in:

[0015] In the formula: This indicates the similarity between region i and region j in their policies related to new energy vehicles. and Let these represent the policy feature vectors of region i and region j, respectively. Indicates the area The L2 norm; Indicates the area The L2 norm.

[0016] The adjacency matrix A of the battery technology diagram is calculated using the following formula. battery Elements in:

[0017] In the formula: This indicates the strength of the association between brand i and battery technology j; This indicates the energy density of the battery used by brand i; This indicates the charge / discharge efficiency of battery technology j; This indicates the cost of using battery technology j in brand i; To prevent small constants with a denominator of zero.

[0018] Step S5: Input historical price time series data, influencing factor data and various adjacency matrices into the trained price prediction model, and output the corresponding used new energy vehicle price prediction data; The specific method is as follows: S51. Data Input: Input historical price time series data, influencing factor data, and various predefined adjacency matrices as training data into the prediction model; S52, Spatiotemporal Attention Weighting: Through a spatiotemporal attention mechanism network, dynamic weights are assigned to the input historical price time series data, influencing factor data, and each adjacency matrix in terms of time and space, generating spatiotemporal weighted time series data corresponding to each adjacency matrix; S53. Temporal Feature Extraction: Using a long short-term memory network, nonlinear temporal modeling is performed on the spatiotemporally weighted temporal data of each adjacency matrix to capture its dynamic change patterns and output the complete time series features of each adjacency matrix.

[0019] S54. Spatial Feature Aggregation: Based on the temporal features of each adjacency matrix, spatial relationship modeling is performed through a graph convolutional network: First, graph structure features are propagated on a single adjacency matrix to generate local spatial features; then, the spatial features of all adjacency matrices are integrated through a feature fusion layer to form a spatiotemporal joint representation.

[0020] S55, High-order feature decoding: Deep feature extraction is performed on the spatiotemporal joint representation, and the predicted value of the price of used new energy vehicles is obtained through mapping by a fully connected layer; S56: Parameter optimization: Calculate the loss function between the predicted price and the actual value, and update the network parameters using the backpropagation algorithm; S57. Repeat S51 to S56 iteratively to train the model until the model converges or reaches the maximum number of iterations.

[0021] Step S6: Use the price prediction data of used new energy vehicles output by the price prediction model as the market price prediction result.

[0022] The beneficial effects of this invention are as follows: By combining the deep learning architecture of LSTM and GCN, it effectively solves the problem of insufficient historical data in the used new energy vehicle market, enabling the extraction of more valuable information from limited historical data; by constructing multiple spatial dependency models, it comprehensively considers various influencing factors such as policies, market demand, production costs, and technological progress, making the prediction results more comprehensive and accurate; by adopting a spatiotemporal attention mechanism, it can adaptively adjust the importance weights of different time points and different spatial relationships, improving the model's adaptability to complex market environments; compared with traditional single time series analysis methods or simple regression models, the method of this invention has a significant improvement in prediction accuracy, providing more reliable price prediction references for all participants in the used new energy vehicle market. Attached Figure Description

[0023] Figure 1 Flowchart of methods for predicting the price of used new energy vehicles; Figure 2 LSTM-GCN model structure diagram; Figure 3 LSTM structure diagram; Figure 4 GCN structure diagram; Figure 5 Schematic diagram of spatiotemporal attention mechanism; Figure 6 Spatiotemporal attention fusion feature architecture diagram. Detailed Implementation

[0024] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] Example: Combination Figure 1 As shown, a method for predicting the price of used new energy vehicles based on the LSTM-GCN model includes the following steps: S1: Obtain historical price time-series data and related influencing factors data for the used new energy vehicle market; S2: Clean and standardize the historical price time series data and related influencing factor data of the used new energy vehicle market; S3: Based on historical price data and related influencing factor data, conduct a structural analysis of the used new energy vehicle market, quantify the relationship between market dimensions and prices, and construct a structured feature input for model training; S4: Based on the market structure, construct multiple models to represent spatial dependencies, and for each spatial dependency, use graph computing methods to generate several corresponding adjacency matrices; S5: Input historical price time series data, influencing factor data and various adjacency matrices into the trained price prediction model, and output the corresponding used new energy vehicle price prediction data; S6: Use the price forecast data of used new energy vehicles output by the price forecast model as the market price forecast result.

[0026] Step S1 specifically includes: In this embodiment, historical price data of major domestic used new energy vehicle brands from January 2018 to December 2023 are obtained, including monthly average sales price, promotional price, and official guide price, etc. Simultaneously, data on relevant influencing factors are collected, including vehicle type, supplier procurement amount, regional economic level, competitor vehicle prices, and market supply and demand data. Specifically, vehicle type includes pure electric sedans, pure electric SUVs, plug-in hybrid sedans, and plug-in hybrid SUVs; supplier procurement amount includes monthly procurement amount data for battery suppliers, motor suppliers, and electronic control system suppliers; regional economic level includes economic indicators such as GDP and per capita disposable income of various provinces and cities; competitor vehicle prices include prices of comparable gasoline vehicles and prices of international brand new energy vehicles in China; market supply and demand data includes monthly production, sales volume, inventory, and order volume.

[0027] Step S2, in this embodiment, involves preprocessing the collected historical price time-series data and related influencing factor data, including organizing and correcting the data. The specific method is as follows: S21: Time series interpolation is used to complete the missing price data, and the isolated forest algorithm is used to check and correct outliers. Finally, the multivariate data is standardized using Min-Max.

[0028] Specifically, for missing values ​​in the price data, time series interpolation is used for imputation. First, missing points in the time series price data for each vehicle model are detected. Then, based on the price data before and after the missing points, methods such as linear interpolation, spline interpolation, or moving average interpolation are used to estimate the missing values. For example, linear interpolation is used for data points missing for no more than 3 consecutive months; cubic spline interpolation is used for data points missing for 3-6 consecutive months; and a seasonally adjusted ARIMA model is used for predictive imputation of data points missing for more than 6 consecutive months.

[0029] For potential outliers, the Isolation Forest algorithm is used for detection. An outlier detection threshold is set to 0.65. For data points with outlier scores exceeding this threshold, a comparison is made with the moving average of the preceding and following three months. If the deviation exceeds 30%, it is considered an outlier and replaced with the median of the preceding and following data.

[0030] Min-Max standardization is performed on all numerical data, mapping each index value to the [0,1] interval. The standardization formula is as follows:

[0031] in, The original data values, This is the minimum value of the indicator. This is the maximum value of the indicator. These are the standardized data values.

[0032] S22: Daily data is aggregated into monthly averages, and quarterly data is decomposed into monthly data through linear interpolation; and vehicle configuration features are processed using Embedding.

[0033] For daily price and sales data, the arithmetic mean of each month is calculated as monthly data; for quarterly economic indicators (such as GDP, consumer confidence index, etc.), linear interpolation is used to decompose them into monthly data.

[0034] For classification features such as vehicle configuration, embedding technology is used. First, a vehicle category dictionary is constructed, assigning a unique ID to each category. Then, an 8-dimensional embedding vector is trained to represent each category. The choice of embedding dimension is determined based on the square root of the number of categories. For the approximately 60 vehicle categories involved in this embodiment, an 8-dimensional embedding space is sufficient to capture the semantic relationships between categories. Embedding training uses a Skip-gram model, trained based on co-occurrence relationships in vehicle sales data, with a learning rate of 0.01 and 100 training epochs.

[0035] Step S3: In this embodiment, a used new energy vehicle market structure model is constructed by performing multi-dimensional analysis on the processed historical price data and related influencing factor data. Specifically, the market structure is analyzed from four dimensions: competitive relationships, supply chain relationships, policy and economic impacts, and power battery technology.

[0036] Competitive Relationship Analysis: Based on factors such as vehicle price range, overlap of target consumer groups, and similarity of features and configurations, this analysis identifies competing vehicle models and brands in the market. The intensity of competition between brands is quantified by calculating indicators such as price correlation coefficients and cross-elasticity of market share changes.

[0037] Supply chain relationship analysis: Based on supplier data, construct a supply network between the brand and key component suppliers. Analyze indicators such as purchase amount, purchase frequency, and dependence level of different brands and suppliers to identify key nodes and potential risk points in the supply chain.

[0038] Policy-Economic Impact Analysis: Based on factors such as the economic development level, consumption capacity, and policy support of various regions, this study analyzes the impact of different regional markets on the price of used new energy vehicles. It also calculates the strength of economic ties between regions and constructs a policy-economic impact network.

[0039] Power Battery Technology Analysis: Based on power battery technology data and supplier relationships, this study constructs a technology supply network between brands and battery technology suppliers. It analyzes key technical elements such as the types of battery technologies used by different brands, performance indicators, and supplier relationships to identify technology evolution paths and supply chain risks.

[0040] Step S4: In this embodiment, based on the market structure analysis results, a model representing spatial dependencies is constructed, and a corresponding adjacency matrix is ​​generated. The specific method is as follows: S41: By constructing graphs, we model the spatial dependencies of market structure in three ways: competitive relationships, supply chain relationships, and policy and economic impacts, resulting in competitive relationship graphs, supply chain relationship graphs, policy and economic impact graphs, and battery technology graphs.

[0041] Competitive relationship graph construction: Brands in the market are represented as nodes in the graph, and the competitive relationships between brands are represented as edges. Node attributes include brand historical price, market share, product line richness, etc.; the edge weights represent the intensity of competition, which are calculated by combining factors such as price similarity and overlap of target customer groups.

[0042] Supply chain relationship graph construction: Brands and major component suppliers are represented as nodes in the graph, and procurement relationships are represented as edges. Node attributes include brand production scale, supplier capacity, etc.; the edge weights represent supply chain strength, calculated from factors such as procurement amount and procurement frequency.

[0043] Policy-Economic Impact Map Construction: Different regional markets are represented as nodes in the graph, and the similarity of policy environments between regions is represented as edges. Node attributes include regional policy characteristics, such as the amount of fiscal subsidies, right-of-way privileges, and the intensity of charging subsidies; the edge weights represent the strength of the policy impact correlation, calculated from the similarity of policy characteristics between regions.

[0044] Battery technology graph construction: Vehicle brands and battery technology types are represented as two types of nodes in the graph, with technology application relationships as edges. Brand node attributes include brand production scale, battery performance requirements, and R&D investment ratio; technology node attributes include technology maturity, energy density, and average cost. The edge weights represent the application intensity between the brand and battery technology, calculated comprehensively based on factors such as energy density, charge / discharge efficiency, application cost, and number of vehicle models.

[0045] S42: Calculate the adjacency matrix A for the competition graph, supply chain graph, policy and economic impact graph, and battery technology graph. comp A supply A policy and A battery .

[0046] Adjacency matrix A of the competition graph compCalculate: For N brands in the market, construct an N×N adjacency matrix A. comp The element C in the matrix ij The competitive intensity between brand i and brand j is represented by the following formula:

[0047] In the formula: C ij P represents the competitive intensity between brand i and brand j; i and P j P represents the average price of brand i and brand j, respectively; max Indicates the highest price in the market; S ij This indicates the degree of overlap between brand i and brand j in the target customer group.

[0048] Specifically, P i and P j Calculated using the average selling price of each brand over the past 12 months, taking into account changes in price trends; P max The highest selling price of all brands in the market is taken; Sij calculates the similarity between two brands by analyzing the degree of overlap in dimensions such as age distribution, income level, and car purchase preferences, with a value range of [0,1].

[0049] Adjacency matrix A of the supply chain relationship graph supply Calculate: For N brands and M major suppliers in the market, construct an N×M adjacency matrix A. supply Element A in the matrix ij The supply chain strength between brand i and supplier j is calculated using the following formula:

[0050] In the formula: A ij Indicates the supply chain strength between brand i and supplier j; T ij T represents the amount that brand i purchases from supplier j; max This represents the maximum single purchase amount for brand i; To prevent small constants with zero denominators, the value is set to 0.0005.

[0051] Specifically, T ij The cumulative purchase amount over the past 12 months is used, taking into account purchase frequency and purchase growth rate; T max Find the maximum purchase amount for brand i among all suppliers.

[0052] Adjacency matrix A of the policy economic impact diagram policy Calculation: For K regional markets, construct a K×K adjacency matrix A policy The element E in the matrix ijThe intensity of the economic impact between region i and region j is calculated using the following formula:

[0053] In the formula: This indicates the similarity between region i and region j in their policies related to new energy vehicles. and Let these represent the policy feature vectors of region i and region j, respectively. Indicates the area The L2 norm; Indicates the area The L2 norm.

[0054] Specifically, the policy feature vector should include key policy indicators that affect the price of used new energy vehicles.

[0055] E=[subsidy,tax_exemption,traffic_privilege,charging_subsidy,purchase_restriction]

[0056] In the formula: subsidy represents the amount of local government subsidies for the purchase of new energy vehicles; tax_exemption represents whether vehicle purchase tax is exempted; traffic_privilege represents whether traffic privileges such as no restrictions on driving or license plates are enjoyed; charging_subsidy represents the level of subsidies or policy support for the construction of charging piles; purchase_restriction represents whether a purchase restriction policy for new energy vehicles is implemented.

[0057] Adjacency matrix A of battery technology diagram battery Calculate: For N car brands and M battery technology types, construct an N×M adjacency matrix A. battery Elements in the matrix The strength of the technological correlation between brand i and battery technology j is calculated using the following formula:

[0058] In the formula: This indicates the strength of the association between brand i and battery technology j; This indicates the energy density of the battery used by brand i; This indicates the charge / discharge efficiency of battery technology j; This indicates the cost of using battery technology j in brand i; To prevent small constants with a denominator of zero.

[0059] Step S5: In this embodiment, a price prediction model based on LSTM-GCN is constructed, and the processed data is input into the model for training and prediction. The specific method is as follows: S51: Data Input. Input historical price time series data, influencing factor data, and various predefined adjacency matrices as training data into the prediction model.

[0060] Specifically, data from January 2018 to December 2022 will be used as the training set, and data from January 2023 to December 2023 will be used as the test set. Model inputs include: Historical price time series data: monthly price series for each brand, in the shape [N, T, 1], where N is the number of brands and T is the time step; Influencing factors data include vehicle type embedding vectors, supplier procurement data, regional economic indicators, competitor prices, market supply and demand data, etc., with a shape of [N, T, F], where F is the feature dimension; Adjacency matrix: includes A comp A supply A policy and A battery .

[0061] S52: Spatiotemporal attention weighting. Through a spatiotemporal attention mechanism network, dynamic weights are assigned to the input historical price time series data, influencing factor data, and each adjacency matrix in terms of time and space, generating spatiotemporally weighted time series data corresponding to each adjacency matrix.

[0062] Spatiotemporal attention is a deep learning technique used to simultaneously model dynamic relationships between temporal and spatial dimensions. Through adaptive weight allocation, it automatically identifies key time points in time series (such as moments of price mutations) and important nodes in spatial networks (such as core suppliers or high-influence regions), thereby more accurately capturing spatiotemporal dependencies in complex systems.

[0063] Spatiotemporal attention mechanisms consist of two parts: temporal attention and spatial attention. Time-based attention mechanism: For each brand's historical price series and influencing factor series, calculate the importance weights at different time steps. Let... Represents the input features at time t, and the temporal attention weights. The calculation is as follows:

[0064] In the formula, It is a learnable vector whose role is to extract key features from the nonlinear transformation of the hidden state to determine which time steps are more important. It is a learnable weight matrix used to map the input sequence to a new feature space. This matrix determines how to combine the input features in order to better capture temporal dependencies. It is a bias term used to adjust the result of the linear transformation, enabling the model to adapt to different data distributions.

[0065] Time-weighted features The calculation is as follows: =

[0066] Spatial attention mechanism: Based on the adjacency matrix, it calculates the spatial dependency weights between different nodes. For a competitive adjacency matrix A... comp Spatial attention weights The calculation is as follows:

[0067] In the formula, and Let i and j represent the feature vectors of node i and node j in the graph, respectively; || represents the feature concatenation operation, which concatenates the feature vectors of the two nodes together to form a longer vector.

[0068] Spatial weighted features The calculation is as follows: =

[0069] For the supply chain relationship adjacency matrix A supply Policy and economic impact adjacency matrix A policy Battery technology diagram A battery It employs a similar spatial attention mechanism for computation.

[0070] S53: Temporal Feature Extraction. Using a Long Short-Term Memory (LSTM) network, nonlinear temporal modeling is performed on the spatiotemporally weighted temporal data of each adjacency matrix to capture its dynamic changes and output the complete time series features of each adjacency matrix.

[0071] Specifically, three LSTM networks are constructed for the spatiotemporally weighted time-series data to process time-series data related to competition, supply chain relationships, and policy and economic impacts, respectively. The structure of each LSTM network is as follows: Input layer: Receives time-series data weighted by time and space; LSTM layer 1: Contains 64 hidden units, uses the tanh activation function, and has a dropout rate of 0.2; LSTM layer 2: Contains 32 hidden units, uses the tanh activation function, and has a dropout rate of 0.2; Output layer: Outputs the hidden state at each time step as the complete time series feature.

[0072] The forward propagation process of an LSTM cell is as follows:

[0073] in, , , These are the input gate, the forget gate, and the output gate, respectively. This is the unit state; The candidate state is generated using tanh to determine the potential update value. It is in a hidden state; is the sigmoid activation function; ⊙ represents element-wise multiplication.

[0074] S54: Spatial Feature Aggregation. Based on the temporal features of each adjacency matrix, spatial relationship modeling is performed through a graph convolutional network (GCN): first, graph structure features are propagated on a single adjacency matrix to generate local spatial features; then, the spatial features of all adjacency matrices are integrated through a feature fusion layer to form a spatiotemporal joint representation.

[0075] Specifically, three GCN networks are constructed for the time-series features output by the LSTM, respectively handling spatial dependencies related to competition, supply chain relationships, and policy-economic impacts. The structure of each GCN network is as follows: Input layer: Receives time series features output by LSTM; GCN layer 1: Contains 32 convolutional kernels, using the ReLU activation function; GCN layer 2: Contains 16 convolutional kernels and uses the ReLU activation function; Output layer: Outputs the aggregated feature matrix.

[0076] The forward propagation process of the GCN layer is as follows:

[0077] in, Let A be the node feature matrix of the l-th layer, and A be the adjacency matrix (A... comp A supply Apolicy Or A battery D is the degree matrix. Let be the weight matrix of the l-th layer. This is the activation function.

[0078] The aggregated feature matrices output by the four GCN networks are fused using an attention mechanism to obtain the spatiotemporal integrated features: +

[0079] in, , , , These are the aggregated feature matrices output by the four GCN networks, respectively. , , , The corresponding attention weights are calculated using learnable parameters.

[0080] S55: High-order feature decoding. Deep feature extraction is performed on the spatiotemporal joint representation, and the predicted price of used new energy vehicles is obtained through mapping using a fully connected layer.

[0081] Specifically, for the fused spatiotemporal integrated features, high-order features are extracted using a fully connected neural network to generate price prediction results: Fully connected layer 1: Contains 64 neurons, uses the ReLU activation function, and has an L2 regularization coefficient of 0.001; Fully connected layer 2: contains 32 neurons, uses the ReLU activation function, and has an L2 regularization coefficient of 0.001; Output layer: Contains neurons for predicting the number of time steps (e.g., 3 neurons for predicting prices over the next 3 months), using a linear activation function.

[0082] The forward propagation process of a fully connected layer is as follows:

[0083] Where Z represents the spatiotemporal integrated features, W and b are learnable parameters, and Y is the predicted output.

[0084] S56: Parameter Optimization. Calculate the loss function between the predicted price and the actual value, and update the network parameters using the backpropagation algorithm.

[0085] Specifically, the mean squared error (MSE) is used as the loss function to calculate the difference between the predicted price and the actual price:

[0086] At the same time, L1 regularization and L2 regularization are introduced to prevent the model from overfitting:

[0087] in, and These are the L1 and L2 regularization coefficients, set to 0.0001 and 0.0005 respectively.

[0088] The Adam optimizer was used for parameter updates. The initial learning rate was set to 0.001, and a learning rate decay strategy was adopted, with the learning rate decaying to 0.9 times the original value every 50 epochs.

[0089] S57: Repeat steps S51 to S56 to iterate the training of the model until the model converges or the maximum number of iterations is reached.

[0090] Specifically, the maximum number of iterations is set to 500 epochs, and the early stopping strategy is to stop training if the loss on the validation set no longer decreases after 20 consecutive epochs. After each epoch, model performance metrics are calculated on the validation set, including mean squared error (MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).

[0091] During training, a mini-batch gradient descent method with a batch size of 64 was used, and the order of training data was randomly shuffled in each epoch. To prevent gradient explosion, a gradient clipping threshold of 5.0 was set.

[0092] Step S6: Use the price prediction data of used new energy vehicles output by the price prediction model as the market price prediction result.

[0093] In this embodiment, the trained LSTM-GCN model is applied to test data to generate a price prediction result for used new energy vehicles for the next three months. Specifically, historical data from October 2023 to December 2023 is used as input to predict the price trend from January 2024 to March 2024.

[0094] The forecast results include monthly expected prices for each brand and model, as well as confidence intervals for price changes (95% confidence level). Simultaneously, based on the forecast results, a price trend report is generated, including overall market price trends, segmented market price trends, and key brand price trends, providing decision support for market participants.

Claims

1. A method for predicting the price of used new energy vehicles based on the LSTM-GCN model, characterized in that, The steps are as follows: Step S1: Obtain historical price time series data and related influencing factor data for the used new energy vehicle market; Step S2: Clean and standardize the historical price time series data and related influencing factor data of the collected used new energy vehicle market; Step S3: Analyze the structure of the used new energy vehicle market based on historical price data and related influencing factors; Step S4: Based on the structure of the used new energy vehicle market, construct a variety of models to represent spatial dependencies, and generate several corresponding adjacency matrices for each spatial dependency using graph computing methods. Step S5: Input historical price time series data, influencing factor data and various adjacency matrices into the trained price prediction model, and output the corresponding used new energy vehicle price prediction data; Step S6: Use the price prediction data of used new energy vehicles output by the price prediction model as the market price prediction result.

2. The method for predicting the price of used new energy vehicles based on the LSTM-GCN model according to claim 1, characterized in that, In step S1, the relevant influencing factors include: vehicle positioning and configuration, supplier procurement amount, policy and economic factors, competitor vehicle prices, power battery technology and market supply and demand data.

3. The method for predicting the price of used new energy vehicles based on the LSTM-GCN model according to claim 1, characterized in that, The specific method in step S2 is as follows: S21. Use time series interpolation to complete missing price data, and use the isolated forest algorithm to check and correct outliers, and perform Min-Max standardization on multivariate data; S22. Aggregate daily data into monthly averages, and decompose quarterly data into monthly data through linear interpolation; and use embedding to process vehicle configuration features.

4. The method for predicting the price of used new energy vehicles based on the LSTM-GCN model according to claim 1, characterized in that, The specific method in step S4 is as follows: S41. By constructing graphs, model the four spatial dependencies of market structure: competitive relationship, supply chain relationship, policy and economic impact, and power battery technology impact, to obtain competitive relationship graph, supply chain relationship graph, policy and economic impact graph, and battery technology graph; S42. Calculate the adjacency matrix A for the competition graph, supply chain graph, policy and economic impact graph, and battery technology graph. comp A supply A policy and A battery .

5. The method for predicting the price of used new energy vehicles based on the LSTM-GCN model according to claim 4, characterized in that, In S42, the adjacency matrix A of the competition graph is calculated using the following formula. comp Elements in: In the formula: C ij P represents the competitive intensity between brand i and brand j; i and P j P represents the average price of brand i and brand j, respectively; max Indicates the highest price in the market; S ij This indicates the degree of overlap between brand i and brand j in the target customer group.

6. The method for predicting the price of used new energy vehicles based on the LSTM-GCN model according to claim 4, characterized in that, In step S42, the adjacency matrix A of the supply chain relationship graph is calculated using the following formula. supply Elements in: In the formula: A ij Indicates the supply chain strength between brand i and supplier j; T ij This represents the amount that brand i purchases from supplier j; T max This represents the maximum single purchase amount for brand i; To prevent small constants with a denominator of zero.

7. The method for predicting the price of used new energy vehicles based on the LSTM-GCN model according to claim 4, characterized in that, In S42, the adjacency matrix A of the policy economic impact diagram is calculated using the following formula. policy Elements in: In the formula: This indicates the similarity between region i and region j in their policies related to new energy vehicles. and Let these represent the policy feature vectors of region i and region j, respectively. Indicates the region The L2 norm; Indicates the region The L2 norm.

8. The method for predicting the price of used new energy vehicles based on the LSTM-GCN model according to claim 4, characterized in that, In step S42, the adjacency matrix A of the battery technology diagram is calculated using the following formula. battery Elements in: In the formula: This indicates the strength of the association between brand i and battery technology j; This indicates the energy density of the battery used by brand i; This indicates the charge / discharge efficiency of battery technology j; This indicates the cost of using battery technology j in brand i; To prevent small constants with a denominator of zero.

9. The method for predicting the price of used new energy vehicles based on the LSTM-GCN model according to claim 1, characterized in that, The specific method in step S5 is as follows: S51. Data Input: Input historical price time series data, influencing factor data, and various predefined adjacency matrices as training data into the prediction model; S52, Spatiotemporal Attention Weighting: Through a spatiotemporal attention mechanism network, dynamic weights are assigned to the input historical price time series data, influencing factor data, and each adjacency matrix in terms of time and space, generating spatiotemporal weighted time series data corresponding to each adjacency matrix; S53. Temporal Feature Extraction: Using a long short-term memory network, nonlinear temporal modeling is performed on the spatiotemporally weighted temporal data of each adjacency matrix to capture its dynamic change patterns and output the complete time series features of each adjacency matrix. S54. Spatial Feature Aggregation: Based on the temporal features of each adjacency matrix, spatial relationship modeling is performed through a graph convolutional network: First, graph structure features are propagated on a single adjacency matrix to generate local spatial features; Then, the spatial features of all adjacency matrices are integrated through a feature fusion layer to form a spatiotemporal joint representation. S55, High-order feature decoding: Deep feature extraction is performed on the spatiotemporal joint representation, and the predicted value of the price of used new energy vehicles is obtained through mapping by a fully connected layer; S56: Parameter optimization: Calculate the loss function between the predicted price and the actual value, and update the network parameters using the backpropagation algorithm; S57. Repeat S51 to S56 iteratively to train the model until the model converges or reaches the maximum number of iterations.

Images