A vehicle energy consumption prediction method and device based on double-branch input and a medium

Abstract

The application discloses a vehicle energy consumption prediction method and device based on double-branch input and a medium, relates to the technical field of vehicle energy consumption prediction, and comprises the following steps: performing multi-modal feature extraction on driving energy consumption data to obtain a normalized time sequence; performing multi-level discrete wavelet transform on the normalized time sequence to determine a standard multi-channel input tensor; extracting local features of the standard multi-channel input tensor to obtain a local space-time feature sequence; calculating and splicing hidden states at each time step of the local space-time feature sequence to determine a hidden state sequence; calculating attention weights corresponding to the hidden state sequence to obtain a context vector; determining an energy consumption feature vector according to a macro working condition feature vector and the context vector; and performing feature aggregation and energy consumption prediction mapping on the energy consumption feature vector to obtain an energy consumption prediction value of a target vehicle. The application solves the technical problems of insufficient robustness and accuracy of vehicle energy consumption prediction in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle energy consumption prediction, and in particular to a vehicle energy consumption prediction method based on double-branch input, a device and a medium. BACKGROUND

[0002] With the increasingly severe global energy situation, reducing vehicle energy consumption has become the core issue of the development of the automobile industry. Accurate energy consumption prediction is a key enabling technology to achieve the goal of energy saving and consumption reduction, and is of great significance for optimizing driving behavior, planning economic routes, and alleviating the range anxiety of new energy vehicle users.

[0003] The vehicle energy consumption prediction technology in the prior art mainly includes: establishing a mathematical model from the longitudinal dynamics equation of the vehicle, regarding energy consumption prediction as a regression problem, predicting by machine learning method, and predicting based on a single deep learning model such as long short-term memory network (LSTM). For the technical problems existing in the above technical means, first, the vehicle energy consumption prediction method of establishing a mathematical model often considers driving resistance, vehicle acceleration dynamics and powertrain efficiency characteristics. The accuracy of the model is highly dependent on accurate vehicle inherent parameters and complex powertrain efficiency data. The calculation accuracy is unstable due to environmental changes. Second, the vehicle energy consumption prediction method based on machine learning learns the nonlinear relationship between energy consumption and various influencing factors from historical data. The existing machine learning model has limited ability to capture the internal dynamics and time sequence dependence of time series data, and it is difficult to distinguish between two macroscopic features with completely different internal driving modes. The upper limit of accuracy is insufficient. Third, compared with traditional machine learning, the vehicle energy consumption prediction method of deep learning model can directly process raw time series and automatically learn time sequence features, which significantly improves the prediction accuracy, but it is less sensitive than convolutional neural network (CNN) in extracting highly localized instantaneous features. Further, although a single CNN can capture local patterns, its receptive field is limited and it is difficult to understand long-term time sequence features. SUMMARY

[0004] The embodiments of the present application provide a vehicle energy consumption prediction method based on double-branch input, a device and a medium, which solve the technical problems of insufficient robustness and accuracy of vehicle energy consumption prediction in the prior art.

[0005] In a first aspect, the embodiments of the present application provide a vehicle energy consumption prediction method based on double-branch input, characterized in that the method comprises: obtaining driving energy consumption data of a real-time working condition of a target vehicle, and performing multi-modal feature extraction on the driving energy consumption data to obtain a normalized time sequence of the driving energy consumption data; performing multi-level discrete wavelet transform on the normalized time sequence to determine a standard multi-channel input tensor; extracting local features of the standard multi-channel input tensor through a preset convolutional neural network to obtain a local spatio-temporal feature sequence; calculating and concatenating hidden states at each time step of the local spatio-temporal feature sequence through a long short-term memory network to determine a hidden state sequence; calculating attention weights corresponding to each alignment score in the hidden state sequence to obtain a context vector corresponding to the hidden state sequence; determining a macro working condition feature vector of the target vehicle based on the normalized time sequence, and determining an energy consumption feature vector of the target vehicle in a driving working condition through vector dimension alignment and Transformer encoding based on the macro working condition feature vector and the context vector; and performing feature aggregation and energy consumption prediction mapping on the energy consumption feature vector to obtain an energy consumption prediction value of the target vehicle.

[0006] In an implementation manner of the present application, the driving energy consumption data comprises: a vehicle speed time sequence corresponding to each independent working condition, and total energy consumption corresponding to each independent working condition; the multi-modal feature extraction on the driving energy consumption data to obtain the normalized time sequence of the driving energy consumption data specifically comprises: screening a global maximum value and a global minimum value corresponding to the vehicle speed time sequence, and calculating normalized values corresponding to each vehicle speed time element in the vehicle speed time sequence based on the global maximum value and the global minimum value to obtain a normalized time sequence input; calculating a vehicle driving feature vector corresponding to the driving energy consumption data, and integrating the vehicle driving feature vector into a macro working condition feature vector; wherein the vehicle driving feature vector comprises: average speed, driving mileage, and speed standard deviation; constructing a sample label corresponding to the total energy consumption, and integrating the normalized time sequence input, the macro working condition feature vector, and the sample label to obtain the normalized time sequence.

[0007] In an implementation manner of the present application, the normalized time sequence is subjected to multi-level discrete wavelet transform to determine the standard multi-channel input tensor, specifically comprising: setting the normalized time sequence as initial approximation coefficients, and calculating first approximation coefficients corresponding to the initial approximation coefficients through low-pass filtering and down-sampling; calculating first detail coefficients corresponding to the initial approximation coefficients through high-pass filtering and down-sampling; iteratively calculating second approximation coefficients and second detail coefficients corresponding to the first approximation coefficients until a preset decomposition level is reached to determine third detail coefficients; obtaining a wavelet coefficient set based on the first detail coefficients, the second detail coefficients and the third detail coefficients; and reconstructing the wavelet coefficient set to determine a reconstructed frequency band component; wherein the reconstructed frequency band component comprises a reconstructed low-frequency component, a reconstructed medium-frequency component and a reconstructed high-frequency component; and the normalized time sequence and the reconstructed frequency band component are spliced to determine the standard multi-channel input tensor.

[0008] In an implementation manner of the present application, a local feature of the standard multi-channel input tensor is extracted through a preset convolutional neural network to obtain a local spatio-temporal feature sequence, specifically comprising: calculating a feature sequence corresponding to a convolution kernel filter of the standard multi-channel input tensor through the preset convolution kernel filter; obtaining a local original feature map corresponding to the standard multi-channel input tensor through sliding window processing based on the feature sequence, and performing normalization processing on the original feature map to obtain a normalized feature map; obtaining an activated feature map through nonlinear transformation according to the normalized feature map; and calculating a pooling value of each sample corresponding to a time step of the activated feature map to obtain the local spatio-temporal feature sequence.

[0009] In an implementation manner of the present application, a hidden state sequence is determined by calculating and splicing hidden states at each time step of the local spatio-temporal feature sequence through a long short-term memory network, specifically comprising: setting an input gate and a candidate cell state corresponding to the local spatio-temporal feature sequence; updating a cell state corresponding to the local spatio-temporal feature sequence based on the input gate and the candidate cell state to determine a current cell state; calculating an output part of the current cell state and performing tanh function normalization processing on the output part to determine a hidden state corresponding to the local spatio-temporal feature sequence; obtaining hidden states corresponding to each time step of the local spatio-temporal feature sequence and splicing the hidden states to determine the hidden state sequence.

[0010] In an implementation manner of the present application, an attention weight corresponding to each alignment score in the hidden state sequence is calculated to obtain a context vector corresponding to the hidden state sequence, specifically comprising: calculating an alignment score corresponding to each time step in the hidden state sequence, and performing normalization processing on the alignment score through a Softmax function to obtain a score matrix corresponding to all time steps; calculating an attention weight of the hidden state sequence based on the score matrix; and performing weighted summation on the hidden state sequence according to the attention weight to obtain the context vector.

[0011] In an implementation form of the present application, according to the macro operating condition feature vector and the context vector, the energy consumption feature vector under the target vehicle operating condition is determined through vector dimension alignment and Transformer coding, specifically comprising: through the same dimension token, the sequence format of the macro operating condition feature vector and the context vector is unified to determine the standard coding vector; wherein the standard coding vector includes: the standardized macro operating condition feature vector, the standardized context vector; the multi-dimensional transpose of the standard coding vector is performed to obtain the tensor split item, and the attention score matrix corresponding to the tensor split item is calculated through the scaling dot product attention; based on the attention score matrix, the recovery tensor corresponding to the standard coding vector is calculated in parallel, and the recovery tensor is integrated to determine the multi-head attention output; the multi-head attention output is processed through the feedforward neural network to determine the energy consumption feature vector under the target vehicle operating condition.

[0012] In an implementation form of the present application, the energy consumption feature vector is aggregated and mapped to the energy consumption prediction to obtain the energy consumption prediction value of the target vehicle, specifically comprising: the energy consumption feature vector is aggregated to obtain the fusion feature vector; the fusion feature vector is linearly mapped to the corresponding energy consumption prediction data to obtain the energy consumption prediction value of the target vehicle.

[0013] In a second aspect, the embodiments of the present application also provide a vehicle energy consumption prediction device based on double-branch input, characterized in that the device comprises: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to implement a vehicle energy consumption prediction method based on double-branch input.

[0014] In a third aspect, the embodiments of the present application also provide a non-volatile computer storage medium for vehicle energy consumption prediction based on double-branch input, which stores computer executable instructions, and the computer executable instructions are executed to implement a vehicle energy consumption prediction method based on double-branch input.

[0015] The embodiment of the application provides a vehicle energy consumption prediction method, device and medium based on double-branch input, global standardization processing is performed on an input complete vehicle speed working condition sequence, so that absolute speed comparability between different working conditions is reserved, a macro statistical feature vector capable of summarizing overall characteristics of the working condition is extracted, time-frequency analysis is performed on the standardized vehicle speed time sequence by adopting a multi-level discrete wavelet transform (DWT), and driving behavior is cooperatively observed from different frequency scales; a preset one-dimensional convolutional neural network and a long short-term memory network (LSTM) are used to effectively capture short-term time sequence evolution rules between the local driving modes; a key dynamic event in a vehicle driving process is captured through an attention mechanism; joint feature representation of a vehicle driving state and vehicle energy consumption is realized by performing information interaction and dynamic weighting on the context time sequence vector and the macro working condition feature vector in the same representation space; finally, a total energy consumption prediction value of a complete working condition corresponding to the vehicle driving is directly mapped and output by regression prediction of the aggregated feature vector, and the technical problems of insufficient robustness and accuracy of vehicle energy consumption prediction in the prior art are solved. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the application without imposing undue limitation thereon. In the drawings:

[0017] Figure 1 A flow chart of a vehicle energy consumption prediction method based on double-branch input is provided for the embodiment of the application;

[0018] Figure 2 A schematic diagram of a vehicle energy consumption prediction method based on double-branch input is provided for the embodiment of the application;

[0019] Figure 3 A probability distribution comparison diagram of a key macro feature is provided for the embodiment of the application; wherein, Figure 3 In (a), a probability density distribution histogram of the average travel speed is shown, (b) is a probability density distribution histogram of the average travel speed, (c) is a probability density distribution histogram of the maximum speed, (d) is a probability density distribution histogram of the maximum acceleration, (e) is a probability density distribution histogram of the travel mileage, and (f) is a probability density distribution histogram of the idle time proportion;

[0020] Figure 4 A speed sequence of a single sample and a four-channel wavelet decomposition result diagram thereof are provided for the embodiment of the application; wherein, Figure 4 In (a), a result diagram of an original working condition channel after wavelet decomposition is shown, (b) is a result diagram of a low-frequency channel after wavelet decomposition, (c) is a result diagram of a medium-frequency channel after wavelet decomposition, and (d) is a result diagram of a high-frequency channel after wavelet decomposition;

[0021] Figure 5 A complete model predicted value and true value scatter plot provided for an embodiment of the present application;

[0022] Figure 6 A loss function descent curve of a training process provided for an embodiment of the present application;

[0023] Figure 7 A macroscopic feature SHAP summary plot of a vehicle working condition provided for an embodiment of the present application;

[0024] Figure 8 A time series and macroscopic feature average contribution degree pie chart provided for an embodiment of the present application;

[0025] Figure 9 A time series heat map of a vehicle working condition provided for an embodiment of the present application; wherein, Figure 9 (a) is an original working condition channel time series heat map, (b) is a low frequency channel time series heat map, (c) is a medium frequency channel time series heat map, and (d) is a high frequency channel time series heat map;

[0026] Figure 10 A working condition duration and time series feature contribution degree correlation analysis scatter plot provided for an embodiment of the present application;

[0027] Figure 11 A time series contribution degree distribution histogram provided for an embodiment of the present application;

[0028] Figure 12 A multi-sample time series contribution degree column chart provided for an embodiment of the present application;

[0029] Figure 13 An internal structure schematic diagram of a vehicle energy consumption prediction device based on double-branch input provided for an embodiment of the present application. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described clearly and completely below in combination with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0031] The embodiment of the application provides a vehicle energy consumption prediction method, device and medium based on double-branch input. Global standardization processing is performed on an input complete vehicle speed working condition sequence to retain absolute speed comparability between different working conditions, a macro statistical feature vector capable of summarizing overall characteristics of the working condition is extracted, multi-level discrete wavelet transform (DWT) is adopted to perform time-frequency analysis on the standardized vehicle speed time sequence, and driving behavior is observed from different frequency scales; a preset one-dimensional convolutional neural network and a long short-term memory network (LSTM) are used to effectively capture short-term time sequence evolution rules between the local driving modes; a key dynamic event in vehicle driving is captured through an attention mechanism; a joint feature representation of a vehicle driving state and vehicle energy consumption is realized by performing information interaction and dynamic weighting on the context time sequence vector and the macro working condition feature vector in the same representation space; finally, a total energy consumption prediction value of the complete working condition corresponding to the vehicle driving is directly mapped and output by regression prediction of the aggregated feature vector, and the technical problems of insufficient robustness and accuracy of vehicle energy consumption prediction in the prior art are solved. The technical solutions provided by the embodiments of the application will be described in detail below with reference to the drawings.

[0032] Figure 1 A vehicle energy consumption prediction method based on double-branch input is provided in the embodiment of the application. As shown in Figure 1 The vehicle energy consumption prediction method based on double-branch input provided by the embodiment of the application specifically comprises the following steps:

[0033] In step 101, driving energy consumption data of a target vehicle real-time working condition is acquired, and multi-modal feature extraction is performed on the driving energy consumption data to obtain a normalized time sequence of the driving energy consumption data.

[0034] As an example, since the original vehicle driving data contains relevant parameters of vehicle driving and corresponding vehicle energy consumption in the driving working condition, in order to enable the driving energy consumption data to represent macro statistical features of overall static attributes of the vehicle working condition, the embodiment of the application performs multi-modal feature extraction on the driving energy consumption data to construct multi-modal composite input data for subsequent hierarchical feature abstraction.

[0035] Specifically, the driving energy consumption data includes: a speed time sequence corresponding to each independent working condition, and total energy consumption corresponding to each independent working condition; multi-modal feature extraction is performed on the driving energy consumption data to obtain a normalized time sequence of the driving energy consumption data, including: screening a global maximum value and a global minimum value corresponding to the speed time sequence, and calculating a normalized value corresponding to each speed time element in the speed time sequence based on the global maximum value and the global minimum value to obtain a normalized time sequence input; a vehicle driving feature vector corresponding to the driving energy consumption data is calculated, and the vehicle driving feature vector is integrated into a macro working condition feature vector; wherein the vehicle driving feature vector includes: average speed, driving distance, and speed standard deviation; a sample label corresponding to the total energy consumption is constructed, and the normalized time sequence input, the macro working condition feature vector, and the sample label are integrated to obtain the normalized time sequence.

[0036] Figure 2 A vehicle energy consumption prediction method based on double-branch input provided by an embodiment of the present application is shown in the schematic diagram.

[0037] Figure 3 A probability distribution comparison diagram of key macro features provided by an embodiment of the present application; wherein, Figure 3 Fig. (a) is a probability density distribution histogram of average travel speed, Fig. (b) is a probability density distribution histogram of average travel speed, Fig. (c) is a probability density distribution histogram of maximum speed, Fig. (d) is a probability density distribution histogram of maximum acceleration, Fig. (e) is a probability density distribution histogram of driving distance, and Fig. (f) is a probability density distribution histogram of idle time proportion.

[0038] In one embodiment, first, based on intelligent transportation systems (ITS) and big data technology, high-precision vehicle terminal (T-BOX) real-time records of new energy vehicle big data monitoring related platforms are obtained, taking a van carrying out logistics distribution tasks in an urban environment as an example, data collection covers typical driving trajectories of corresponding vehicle models in a preset research area, and corresponding urban driving cycles (Driving Cycles) include but are not limited to: intersection start-stop caused by traffic signals or congestion, common congestion following in urban peak hours, and smooth cruising on urban expressways.

[0039] Then, the training or prediction input is a complete, actual driving vehicle working condition and energy consumption data set, each independent working condition sample contains a complete speed time sequence , wherein is the total duration of the working condition. Also contains the total energy consumption corresponding to the working condition .

[0040] All training working conditions (assuming a total of ) are traversed, and the global minimum value of all speed data points is found and global maximum The global normalization scale calculation is explained by the following formula.

[0041] (1)

[0042] (2)

[0043] Through single sample processing and feature extraction, each working condition sample is normalized using a global scale for each point in to obtain a normalized time series input , which is explained by the following formula.

[0044] (3)

[0045] Further, the macro working condition feature vector is calculated, specifically including: average speed, driving mileage and speed standard deviation, and the macro feature vector will be an independent token (Token) participating in the subsequent fusion process.

[0046] The average speed is calculated as , the driving mileage is , and the speed standard deviation is .

[0047] Finally, all calculated statistics (i.e. the part corresponding to the macro working condition feature vector) are spliced into a vector and a label is constructed to obtain a normalized time series sequence of driving energy consumption data composed of , and .

[0048] Further, for the statistics of the driving energy consumption data of the target vehicle real-time working condition, in order to capture the inherent statistical law of the real driving behavior, a random process modeling method based on Markov Chain is adopted to generate a large number of synthetic driving working conditions highly similar to the original data in statistical characteristics.

[0049] All state transition frequencies between adjacent time steps are counted second by second based on all preprocessed real driving data. Based on this frequency statistical result, the conditional probability transitions to state is calculated by maximum likelihood estimation method All these probabilities together form an N x N state transition probability matrix P (where N is the total number of discrete states), which is a mathematical abstraction and quantitative description of the real driving behavior patterns.

[0050] After the state transition probability matrix is constructed, the random generation of new driving cycles can be performed, starting from a specified initial state (usually the stationary state) and iteratively sampling according to the state transition probability matrix. At each time step, the next state is randomly determined according to the probability distribution corresponding to the current state in the matrix, thereby generating a state ID sequence with a specific length.

[0051] The generated state ID sequence is mapped back to a continuous speed-acceleration time sequence. Then, post-processing techniques such as smoothing filters are applied to ensure that the generated cycle curve is physically reasonable and smooth, especially to ensure that it can naturally end in a parking state.

[0052] To ensure that the synthesized cycles have practical application value, a quality control mechanism is used, which immediately evaluates the key macro features of each generated cycle, such as total travel distance. A minimum distance threshold is set, and any cycle below this threshold is considered invalid and discarded. This screening mechanism ensures that the final retained cycles meet the basic quality requirements by repeatedly executing the generation-screening cycle.

[0053] The original dataset is finally effectively expanded to include 12000 verified valid driving cycles. For each cycle in the original and augmented datasets, a set of macro feature indicators covering speed, acceleration, travel stage, power, and other dimensions is calculated. Then, a group of representative key features is selected, and the probability distributions of the two datasets are directly compared through visualization. Figure 2 It can be seen that, whether it is average travel speed, maximum speed, travel distance, or idle time ratio, the probability distribution curve of the augmented data (orange) in shape, peak position, and data distribution range is highly overlapped and fitted with the original data (blue). This consistency in distribution indicates that the data augmentation method based on Markov chains successfully captures and reproduces the inherent statistical rules of real driving behavior.

[0054] Step 102, performing a multi-level discrete wavelet transform on the normalized time series to determine a standard multi-channel input tensor.

[0055] Exemplarily, the application adopts a multi-level discrete wavelet transform (DWT) to perform time-frequency analysis on the normalized vehicle speed time series, decomposes and reconstructs the original one-dimensional signal into multiple parallel and equal-length feature channels, realizes the differentiation of components under different working conditions of the vehicle, enables the model to cooperatively examine the driving behavior from different scales, and provides a multi-dimensional data basis for the subsequent CNN to capture cross-channel cooperative patterns.

[0056] Specifically, the normalized time series is subjected to multi-level discrete wavelet transform to determine a standard multi-channel input tensor, including: setting the normalized time series as initial approximation coefficients, and calculating first approximation coefficients corresponding to the initial approximation coefficients through low-pass filtering and downsampling; calculating first detail coefficients corresponding to the initial approximation coefficients through high-pass filtering and downsampling; iteratively calculating second approximation coefficients and second detail coefficients corresponding to the first approximation coefficients until a preset decomposition level is reached to determine third detail coefficients; obtaining a wavelet coefficient set based on the first detail coefficients, the second detail coefficients and the third detail coefficients; and reconstructing the wavelet coefficient set to determine a reconstructed frequency band component; wherein the reconstructed frequency band component includes a reconstructed low-frequency component, a reconstructed medium-frequency component and a reconstructed high-frequency component; and splicing the normalized time series and the reconstructed frequency band component to determine the standard multi-channel input tensor.

[0057] Figure 4 A speed sequence of a single sample and a four-channel wavelet decomposition result thereof are provided for the embodiments of the application; wherein, Figure 4 Fig. (a) is a result graph of an original working condition channel after wavelet decomposition, Fig. (b) is a result graph of a low-frequency channel after wavelet decomposition, Fig. (c) is a result graph of a medium-frequency channel after wavelet decomposition, and Fig. (d) is a result graph of a high-frequency channel after wavelet decomposition.

[0058] The four-channel result obtained by wavelet decomposition of the No. 1600 sample will be used as the input of the CNN-LSTM-Transformer time series processing branch in the model, so that the vehicle dynamic characteristics can be learned more comprehensively.

[0059] In one embodiment, a single working condition sample The generated normalized time series is denoted as To simplify the symbols in the subsequent formulas, the superscript is temporarily omitted and the input sequence is denoted as , and the length of .

[0060] First, through an L-level multi-level wavelet decomposition process, the input signal The decomposition is recursively applied to a series of wavelet coefficients representing different frequency bands. Then, through a selective reconstruction process, using specific subsets of these coefficients, a full time-domain signal (i.e. low, medium and high frequency components) representing a specific frequency band is generated via an inverse discrete wavelet transform (IDWT) that is equal in length to the original signal.

[0061] The speed sequence of each driving condition (uniformly padded or truncated to 2300 time steps) is decomposed into four independent channels:

[0062] (a) Channel One (Original Sequence): The original speed sequence after standardization, preserving the most direct driving information;

[0063] (b) Channel Two (Low Frequency Component): Represents the main trends and long-period changes during driving, such as cruising phases;

[0064] (c) Channel Three (Medium Frequency Component): Captures the regular acceleration and deceleration processes and speed fluctuations;

[0065] (d) Channel Four (High Frequency Component): Reflects the most intense instantaneous changes during driving, such as emergency braking or sudden acceleration.

[0066] From the input signal a set of wavelet coefficients is calculated that characterizes its multi-scale properties, initializing the approximation coefficients of level 0 as the original input signal itself, i.e., .

[0067] For each level of decomposition (from 1 to L), the approximation coefficients of the previous level are used to calculate the approximation coefficients and the detail coefficients of the current level by applying a pair of analysis filters determined by the selected wavelet basis, i.e. by convolving and down-sampling with the low-pass filter and the high-pass filter . where

[0068] is the level index of the decomposition, ; is the element of the approximation coefficient sequence of the previous level; is the discrete time index of the input sequence; is the discrete time index of the output coefficient sequence; and are the coefficients of the low-pass and high-pass filters, respectively, whose calculation is explained by the following formula. where

[0069] (4)

[0070] (5)

[0071] where, Approximation information is extracted by low-pass filtering and down-sampling; Detail information is extracted by high-pass filtering and down-sampling.

[0072] The iteration process is carried out step by step, and each level of decomposition reduces the length of the approximation coefficient sequence by about half until the preset decomposition level L is reached.

[0073] After L-level decomposition, a complete set of wavelet coefficients is obtained, that is, the wavelet coefficient set is denoted as .

[0074] where, represents the highest frequency detail in the signal, represents the relatively low frequency detail, and represents the lowest frequency macro trend in the signal.

[0075] Then, according to the wavelet coefficient set, a complete time-domain signal representing a specific frequency band is generated through inverse discrete wavelet transform (IDWT) which is equal in length to the original signal. IDWT (inverse discrete wavelet transform) is the mathematical inverse operation of DWT, and its core operation unit includes up-sampling and convolution with a pair of reconstruction filters. Up-sampling (Up sampling) aims to restore a shortened coefficient sequence to its previous length. For an input sequence of length N, the output sequence of up-sampling has a length of 2N, which is defined as inserting a zero value between every two adjacent elements of the original sequence, which is explained by the following formula.

[0076] (6)

[0077] To achieve perfect reconstruction of the signal, the reconstruction filter used in inverse transform must form an orthogonal mirror filter pair with the analysis filter used in decomposition. For an analysis low-pass filter of length N , its corresponding reconstruction filter is .

[0078] Based on the above definition, the approximation of the first level is reconstructed from the coefficients of the first level, which is explained by the following formula.

[0079] (7)

[0080] By constructing a specific set of input coefficients (i.e., selectively selecting the set) (Some coefficient vectors in the matrix are set to zero), and then the complete L-level IDWT process is applied to generate the individual target components:

[0081] Reconstructing low-frequency components This represents the overall stable trend of the operating conditions; among them, This represents a zero vector of the same length as the detail coefficients at the corresponding level.

[0082] Reconstructing intermediate frequency components Representing mild acceleration and deceleration behavior under operating conditions, the intermediate frequency in this application is defined as the set of all detail components except for the highest and lowest frequencies.

[0083] Reconstructing high-frequency components It primarily captures aggressive driving behaviors such as rapid acceleration and sudden braking.

[0084] Finally, the original normalized signal and the reconstructed three frequency band components are concatenated along the feature dimension to form a multi-channel input. ;in, The `()` function is used to concatenate multiple arrays to obtain a multi-channel input tensor. Before batch processing, the lengths of different samples will be adjusted by padding or truncation. Unified as a standardized maximum length To obtain the final output To determine the standard multichannel input tensor.

[0085] Step 103: Extract local features of the standard multi-channel input tensor through a pre-defined convolutional neural network to obtain a local spatiotemporal feature sequence.

[0086] For example, the temporal feature processing branch is a key step in capturing vehicle dynamics in this application. Considering the multi-scale characteristics of driving behavior, which has both local abrupt changes (such as sudden braking) and long-term dependencies (such as the persistence of congested road sections), this application designs a cascaded architecture in this branch, sequentially connecting CNN, LSTM and Transformer modules. This architecture aims to deconstruct deep features from microscopic operations to macroscopic trends layer by layer from the four-channel velocity sequence enhanced by wavelet transform. For the convolutional neural network (CNN) part, this application extracts local features from the standard multi-channel input tensor to obtain local spatiotemporal feature sequences, realizing high-level local features representing basic driving, and providing basic feature units for subsequent LSTM temporal analysis.

[0087] Specifically, local spatio-temporal feature sequences are obtained by extracting local features of a standard multi-channel input tensor through a preset convolutional neural network, including: calculating a feature sequence corresponding to a preset convolution kernel filter through the standard multi-channel input tensor; based on the feature sequence, obtaining a local original feature map corresponding to the standard multi-channel input tensor through sliding window processing, and performing normalization processing on the original feature map to obtain a normalized feature map; according to the normalized feature map, an activated feature map is obtained through nonlinear transformation; and a pooling value of each sample corresponding to a time step of the activated feature map is calculated to obtain the local spatio-temporal feature sequence.

[0088] In one embodiment, the convolutional neural network architecture of the present application includes a one-dimensional convolutional layer, a batch normalization layer, a nonlinear activation layer, a maximum pooling layer, etc.

[0089] The input of the convolutional neural network is a standard multi-channel input tensor, which can be denoted as ; wherein, is the batch size, indicating the number of independent working condition samples contained in one processing.

[0090] The one-dimensional convolutional layer performs a sliding window type weighted summation operation on the time dimension of the input sequence through a set of learnable convolution kernel filters to capture different types of local patterns.

[0091] For the th sample in the batch, the value of the th time step, the th channel of the output feature map is calculated through the following formula.

[0092] (8)

[0093] The formula indicates that the th filter is overlaid on the fragment of the input sample starting from the time step with a length of .

[0094] Then, the filter weights are multiplied element by element with the multi-channel input values at the corresponding positions, and all the products are added, and finally the bias term of the filter is added to obtain a single output value. The process slides in the time dimension to generate a new feature sequence for each filter.

[0095] ​Finally, each learnable filter acts as a feature detector, whose weights define the specific local pattern it is looking for. The filter weights are matched point by point with a fixed-length segment of the input sequence and summed up to obtain a scalar score, which reflects the degree of match between the input segment and the pattern.

[0096] To capture the occurrence of a feature across the entire time axis, a new sequence is generated for each filter by traversing every possible position in the input sequence, recording the activation strength of the feature at all positions in the time period, resulting in a feature response map.

[0097] The batch normalization layer normalizes the raw feature map output by the convolutional layer to stabilize the distribution of inputs to each layer, thereby accelerating model convergence and improving training stability.

[0098] For each filter channel of the feature map , first calculate its batch statistics, which are explained by the following formula.

[0099] (9)

[0100] (10)

[0101] Then, normalize, scale and shift each value in the channel, which is explained by the following formula.

[0102] (11)

[0103] (12)

[0104] where is the mean of the th channel of the feature map over the entire batch;

[0105] is the variance of the th channel of the feature map over the entire batch.

[0106] is a very small positive number to prevent the denominator from being zero and increase numerical stability;

[0107] and are the learnable scaling and shifting parameters for the th channel, so that the network can restore the original representation of the data;

[0108] ​ The normalized feature map output by the current layer.

[0109] The nonlinear activation layer performs nonlinear transformation on the normalized feature map to introduce nonlinear representation capability, so that the model can learn more complex function relationships. The application preferably adopts a rectified linear unit (ReLU), and the calculation result is .

[0110] The function judges each element of the input. If the element value is greater than 0, it remains unchanged; if it is less than or equal to 0, it is set to 0. The activation feature map output by the current layer is .

[0111] The max-pooling layer realizes down-sampling by performing a sliding window on the activation feature map and taking the maximum value in the window. The purpose is to reduce the sequence length, reduce the calculation amount, and extract the most significant features in the local area, while giving the model certain translation invariance.

[0112] For each sample and each channel of the activation feature map , the value of the th time step after pooling , the calculation process is explained by the following formula.

[0113] (13)

[0114] wherein, is the size of the pooling window;

[0115] is the stride, that is, the step length of the sliding window each time;

[0116] is the feature sequence finally output by the current layer and the entire convolution block, and the length of the feature sequence is .

[0117] After processing by one or more of the above convolution blocks, the CNN module finally outputs a shorter and deeper local spatiotemporal feature sequence Each time step of the local spatiotemporal feature sequence extracts the local spatiotemporal pattern information of the original input signal in the corresponding time region.

[0118] Step 104, calculate and splice the hidden state of each time step of the local spatiotemporal feature sequence by a long short-term memory network to determine a hidden state sequence.

[0119] Exemplarily, after the CNN completes the extraction of the local features, the application models the timing logic of the abstracted features through a long short-term memory network (LSTM). The LSTM utilizes its unique gating mechanism to capture the time sequence, context dependence and long-term evolution law among the discrete local features.

[0120] Specifically, through the long short-term memory network, the hidden state sequence is determined by calculating and splicing the hidden state at each time step of the local spatio-temporal feature sequence, including: setting the input gate and the candidate cell state corresponding to the local spatio-temporal feature sequence; updating the cell state corresponding to the local spatio-temporal feature sequence based on the input gate and the candidate cell state to determine the current cell state; calculating the output part of the current cell state, and performing tanh function normalization processing on the output part to determine the hidden state corresponding to the local spatio-temporal feature sequence; obtaining the hidden state corresponding to each time step of the local spatio-temporal feature sequence, and splicing the hidden states to determine the hidden state sequence.

[0121] In one embodiment, the local spatio-temporal feature sequence can be denoted as . To simplify the symbols in the subsequent formulas, the input sequence is denoted as ; wherein the sequence length is , and the input feature dimension is ; represents the input feature vector of the sequence at time step (where ).

[0122] At each time step , the LSTM unit receives three inputs: the input feature vector at the current time. The hidden state (Hidden State) at the last time . The cell state (Cell State) at the last time . It selectively updates information and produces output through three key gating units inside: the forget gate, the input gate and the output gate.

[0123] First, define as the number of hidden units of the LSTM, which defines the dimensions of the hidden state and the cell state ;

[0124] respectively acting on four different weight matrices of the input ;

[0125] respectively acting on four different weight matrices of the hidden state at the last time;

[0126] are the bias vectors corresponding to the four computing units, respectively;

[0127] is the Sigmoid activation function;

[0128] is the hyperbolic tangent activation function;

[0129] is the Hadamard product.

[0130] Then, the input gate and the candidate cell state determine which new information will be stored in the cell state.

[0131] Input gate determines which values will be updated and which dimensions need to write new information. The calculation process is explained by the following formula.

[0132] (14)

[0133] Candidate cell state creates a vector containing new candidate values that may be added to the cell state.

[0134] (15)

[0135] Combines the old cell state and the new candidate information to generate the cell state at the current time .

[0136] (16)

[0137] Finally, the output gate determines which part of the cell state will be output. The calculation process is explained by the following formula.

[0138] (17)

[0139] Hidden state update , the updated cell state is normalized by function, and then multiplied by the output of the output gate to obtain the final hidden state. The calculation formula is:

[0140] (18)

[0141] will be executed in turn at each time step of the input sequence , and the output is the hidden state sequence obtained by splicing the hidden states of all time steps , which can be denoted as ; wherein, .

[0142] Step 105, calculate the attention weight corresponding to each alignment score in the hidden state sequence to obtain the context vector corresponding to the hidden state sequence.

[0143] For example, to solve the problem of information redundancy in long-term working conditions, the application introduces an attention mechanism to focus on key information on the basis of the complete driving event sequence constructed by LSTM. By simulating the focus allocation in the human cognitive process, the importance of each time in the sequence to the final energy consumption prediction is dynamically identified and quantified through learnable weights, and finally the variable-length sequence is weighted and compressed into a fixed-length context vector, which is a highly condensed summary of all key dynamic events in the entire vehicle working condition.

[0144] Specifically, calculating the attention weight corresponding to each alignment score in the hidden state sequence to obtain the context vector corresponding to the hidden state sequence includes: calculating the alignment score corresponding to each time step in the hidden state sequence, and normalizing the alignment score through the Softmax function to obtain the score matrix corresponding to all time steps; based on the score matrix, the attention weight of the hidden state sequence is calculated; according to the attention weight, the hidden state sequence is weighted and summed to obtain the context vector.

[0145] In one embodiment, the hidden state sequence can be denoted as for each time step in the input sequence An original, non-normalized importance score is calculated. This score is based on the hidden state vector of each time step which is calculated by itself.

[0146] First, through a small feedforward neural network, each is mapped to an intermediate representation , using as a nonlinear activation function to increase the expression capacity, and the calculation process is explained by the following formula.

[0147] (19)

[0148] where, is a learnable weight matrix used to linearly transform the hidden state vector to an intermediate attention representation space; is the dimension of this space, which is an adjustable hyperparameter; is the corresponding bias vector.

[0149] Then, the intermediate representation is combined with the learnable vector The dot product is performed, which is reduced to a scalar score The calculation process is explained by the following formula.

[0150] (20)

[0151] where, The intermediate attention representation of the time step;

[0152] is a learnable weight vector that projects the intermediate representation to a scalar score;

[0153] is the alignment score of the time step.

[0154] Finally, the alignment scores are converted to a normalized probability distribution by the Softmax function to ensure that the sum of the weights of all time steps is 1, so that it has a clear physical interpretation.

[0155] The scores for all time steps are applied to the Softmax function in the time dimension, and the calculation process is explained by the following formula.

[0156] (21)

[0157] where, is the final attention weight of the time step ; is the score matrix concatenated by the alignment scores of all time steps.

[0158] The original hidden state sequence is weighted and summed using the calculated attention weights to generate a single, fixed-length context vector , and the calculation process is explained by the following formula.

[0159] (22)

[0160] The context vector highly condensed with key information of the sequence is output, and its data form is .

[0161] Step 106, based on the normalized time sequence, determine the macro working condition feature vector of the target vehicle, and according to the macro working condition feature vector and the context vector, determine the energy consumption feature vector of the target vehicle in the driving condition through vector dimension alignment and Transformer coding.

[0162] For example, by using the self-attention mechanism of Transformer, the context temporal vector and the macroscopic condition feature vector are deeply interacted in the same representation space to learn the nonlinear dependency between them, thereby realizing the fusion of two heterogeneous information types, namely driving dynamics and macroscopic conditions, and improving the quality of multimodal information fusion.

[0163] Specifically, based on the macroscopic operating condition feature vector and the context vector, the energy consumption feature vector under the target vehicle's driving conditions is determined through vector dimension alignment and Transformer encoding. This includes: unifying the sequence format of the macroscopic operating condition feature vector and the context vector using tokens of the same dimension to determine the standard encoding vector; wherein the standard encoding vector includes: a standardized macroscopic operating condition feature vector and a standardized context vector; performing multi-dimensional transpose on the standard encoding vector to obtain tensor decomposition terms, and calculating the attention score matrix corresponding to the tensor decomposition terms through scaled dot product attention; based on the attention score matrix, calculating the recovery tensor corresponding to the standard encoding vector in parallel, and integrating the recovery tensors to determine the multi-head attention output; and processing the multi-head attention output using a feedforward neural network to determine the energy consumption feature vector under the target vehicle's driving conditions.

[0164] In one embodiment, the dual-branch data includes: a time-series dynamic feature vector. That is, the context vector output by the attention mechanism. Macroscopic statistical eigenvectors , shape is ;in, It refers to the quantity of macroscopic characteristics.

[0165] Before the aforementioned dual-branch data is fed into the Transformer, features from different sources and with different dimensions are converted into a sequence in a unified format, ensuring that all tokens have the same dimensions. Time-series dynamic tokens Directly by It is obtained because its dimensions already meet the requirements. .

[0166] Macro feature token Through a dedicated linear layer, from Dimensional projection to Dimensions, to determine .

[0167] in, This refers to the model dimension. To maintain consistency, it is set... ;

[0168] Linear layer weight matrix for projecting macro feature

[0169] is the corresponding bias vector

[0170] is the timing dynamic token

[0171] is the macro feature token.

[0172] Further, the above short sequence containing two kinds of information is deeply cross-relation modeled. The input record X In each layer, the multi-head self-attention mechanism calculates the mutual dependence between two tokens, allowing them to dynamically weight and update each other. Subsequently, the position feed-forward network (FFN) will perform a deep nonlinear transformation on each updated token. The entire process is assisted by residual connections and layer normalization to ensure stable and efficient training.

[0173] First, the multi-head self-attention link is performed. The input query Q, key K, and value V tensors (Q, K, V are tensors with shape [B, 2, ] are first split into H heads. By reshaping and transposing operations on the last dimension , the dimension is split into H and two dimensions, and the shape of Q becomes [B, 2, H, ].

[0174] Then, the 1st and 2nd dimensions (sequence length and head number) are exchanged so that the data of all heads are put together and processed in parallel. The final shape of Q becomes [B, 2, H, ]. K and V also undergo exactly the same transformation. This step physically separates the data of H heads in memory, laying the foundation for efficient parallel computation. Among them, is the number of heads, a preset hyperparameter. is the dimension of each head, which is calculated as .

[0175] For the split H heads, independent and parallel scaled dot product attention calculations are performed. For each head , the attention score is calculated as . Scaling and normalization are performed to obtain the final attention weight , and the value is weighted and summed .

[0176] Among them, is the first query, key and value tensors; is the original attention score matrix computed by the head.

[0177] Further, the outputs computed by all H heads in parallel are concatenated along the last feature dimension, recovering the original dimensionality of to a tensor ConcatHeads of shape [B, 2, d_model], and the information is fused by a final linear layer, explained by the following equation.

[0178] (23)

[0179] where is the final output weight matrix of the sublayer, used to fuse the information of all heads. The matrix WOis used to learn how to best linearly combine the independent attention outputs from different subspaces.

[0180] Finally, the output of multi-head attention is added back to its original input, and then layer normalized to stabilize the training process, computed by the following equation. The LayerNorm function will normalize each token of each sample in the batch independently.

[0181] The positional feed-forward network sublayer further performs a non-linear transformation on the output of the self-attention layer to deeply mine and fuse the correlations of each token within the feature dimension, explained in detail by the following equation.

[0182] (24)

[0183] The FFN acts on each token in the sequence independently; where, is the weight and bias of the first linear layer of the FFN;

[0184] is the intermediate layer dimension, which is usually chosen as ;

[0185] () is the activation function;

[0186] is the weight and bias of the second linear layer.

[0187] The output of the FFN is computed as .

[0188] Step 107, performing feature aggregation and energy consumption prediction mapping on the energy consumption feature vector to obtain the energy consumption prediction value of the target vehicle.

[0189] Exemplarily, the application realizes the fusion strategy of the self-attention mechanism based on the Transformer by aggregating the multi-modal features fused through the Transformer into a single representation and mapping it to the final energy consumption prediction value, considering the two types of information as independent tokens for deep dynamic interaction, thereby improving the accuracy of vehicle energy consumption prediction and the robustness of the system in dealing with complex vehicle driving conditions.

[0190] Specifically, the feature aggregation and energy consumption prediction mapping are performed on the energy consumption feature vector to obtain the energy consumption prediction value of the target vehicle, specifically including: performing feature aggregation on the energy consumption feature vector to obtain a fused feature vector; linearly mapping the fused feature vector to corresponding energy consumption prediction data to obtain the energy consumption prediction value of the target vehicle.

[0191] In one embodiment, a simple linear layer is used to linearly map the high-dimensional fused feature vector to a single energy consumption prediction value, which is explained in detail by the following formula.

[0192] (25)

[0193] For the b-th sample in the batch:

[0194] (26)

[0195] wherein, is the weight matrix of the regression head; is the bias term of the regression head; is the final energy consumption prediction value.

[0196] Through the above pre-fusion architecture, the model can utilize the features after deep interaction in the last stage of prediction, thereby meeting the high prediction accuracy.

[0197] Figure 5 A scatter plot of the complete model prediction value and the real value provided by the embodiment of the application.

[0198] The complete model exhibits excellent prediction accuracy on an independent test set containing 1800 samples, with a root mean square error (RMSE) as low as 0.053 and a coefficient of determination (R²) as high as 0.949. The blue scatter points are tightly clustered around the y=x red dashed line representing the ideal case, showing a high degree of linear correlation. This indicates that the model's predicted values are highly consistent with the true values, whether in the low or high energy consumption interval, and the model can achieve accurate fitting. The R² value of up to 0.949 also proves from the data that the model can explain 94.9% of the variability in the energy consumption data in the test set, further confirming its strong nonlinear mapping ability and excellent generalization performance.

[0199] Figure 6 A loss function descent curve of a training process provided for an embodiment of the present application.

[0200] The training set RMSE (blue curve) and the validation set RMSE (red curve) both rapidly decrease at the beginning of training and quickly converge to a very low level after about 2000 iterations. Throughout the training process, the two curves remain close to each other, and there is no phenomenon of the validation set error being significantly higher than the training set error or rebounding. This indicates that the model has been fully and effectively trained during the learning process and has successfully avoided overfitting, ensuring its stable performance on unseen data.

[0201] Figure 7 A macro feature SHAP summary graph of a vehicle working condition provided for an embodiment of the present application.

[0202] In order to open the "black box" of the model and understand the contribution pattern of different input features to the final energy consumption prediction, the present application uses the SHAP method to perform global explainability analysis on the entire test set.

[0203] The features are sorted from top to bottom according to their average SHAP absolute value, and the average acceleration (ms^2) in the acceleration segment, the deceleration time percentage (%) and the idling time percentage (%) are the most important macro features affecting energy consumption prediction.

[0204] Further, the average acceleration in the acceleration segment: The SHAP value distribution range of this feature is the widest, indicating that it has the greatest influence. Among them, samples with higher feature values (red points) almost all correspond to positive SHAP values, which means that more intense acceleration behavior is the primary factor leading to increased energy consumption per 100 kilometers.

[0205] Deceleration / idling time percentage: These two features show similar patterns, i.e., when the feature value is high (red point), the SHAP value is also generally positive. This indicates that longer deceleration and idling times during driving (which usually means frequent start-stop and inefficient energy utilization) will significantly push up the energy consumption per 100 kilometers.

[0206] Speed interval feature: Most of the speed interval proportion features located in the lower half of the figure have low importance, and the SHAP values are concentrated around zero, indicating that the model considers that simply being in a certain high-speed interval itself is less important than the dynamic changes (acceleration and deceleration) and inefficient working conditions (idling) during driving in affecting the 100 km fuel consumption.

[0207] Figure 8 A timing and macro feature average contribution pie chart is provided for the embodiments of the present application.

[0208] The timing and macro feature average contribution pie chart of the present application quantifies the average contribution of the timing feature and the macro feature to the model prediction. The average contribution of the macro feature accounts for as high as 71.2%, while the contribution of the timing feature accounts for 28.8%. This result shows that for the highly generalized index of 100 km fuel consumption, the macro feature describing the statistical characteristics of the entire trip (such as average acceleration, idling proportion, etc.) is the dominant factor in determining the prediction result. Therefore, the instantaneous dynamic information captured by the timing feature still has a high supplementary and correction function, contributing nearly 30% of the prediction weight.

[0209] Figure 9 A vehicle working condition timing heat map is provided for the embodiments of the present application; wherein, Figure 9 (a) is the original working condition channel timing heat map, (b) is the low frequency channel timing heat map, (c) is the medium frequency channel timing heat map, and (d) is the high frequency channel timing heat map.

[0210] In the entire driving process, the timing heat map of the vehicle working condition reflects which time point of the driving behavior contributes most to the final prediction. In particular, in the original channel and the medium frequency channel, in the intervals of about 100-300 seconds, 800-1000 seconds and 1800-2000 seconds, the speed curve fluctuates sharply (frequent acceleration and deceleration), and the corresponding SHAP value also presents significant red (positive contribution) or dark blue (negative contribution). This shows that the model successfully focuses attention on these high-dynamic driving segments that have the greatest impact on energy consumption.

[0211] Figure 10 A working condition duration and timing feature contribution correlation analysis scatter plot is provided for the embodiments of the present application.

[0212] To further explore whether the model has the "intelligent" behavior of dynamically adjusting the decision basis according to the working condition characteristics, the application analyzes the relationship between the working condition duration and the dependence of the model on the time sequence characteristics. There is a significant and moderately strong positive correlation between the working condition duration and the contribution degree of the time sequence characteristics (Pearson correlation coefficient R = 0.625, p-value = 6.219e-196). The p value is much smaller than the threshold of 0.05, indicating that the correlation is statistically extremely significant. That is, the model corresponding to the architecture adopted by the application is not a static mapping function, but can dynamically adjust its decision basis. For simple working conditions with short duration, the model relies more on macro statistics to make quick judgments; and for working conditions with long duration and more complex processes, the model will rely more on detailed time sequence dynamic information learned from the driving process to make more refined predictions.

[0213] Figure 11 A time sequence contribution degree distribution histogram is provided for an embodiment of the application.

[0214] Figure 12 A multi-sample time sequence contribution degree histogram is provided for an embodiment of the application.

[0215] Further, the time sequence contribution degree shows a nearly normal wide distribution in all samples, rather than a fixed value. This adaptive decision-making mode is a key indicator of whether a deep learning model is truly "intelligent", representing the reliability and credibility of the model in practical applications.

[0216] The above is a method embodiment proposed by the application. Based on the same inventive concept, the embodiment of the application also provides a vehicle energy consumption prediction device based on double-branch input, the structure of which is as shown in Figure 13 .

[0217] Figure 13 A schematic diagram of the internal structure of a vehicle energy consumption prediction device based on double-branch input is provided for an embodiment of the application. As shown in Figure 13 , the device includes:

[0218] at least one processor 1301;

[0219] and a memory 1302 in communication connection with the at least one processor;

[0220] The memory 1302 stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor 1301 to enable the at least one processor 1301 to implement a vehicle energy consumption prediction method based on double-branch input.

[0221] Some embodiments of the application provide a vehicle energy consumption prediction method based on double-branch input. Figure 1The application discloses a nonvolatile computer storage medium for vehicle energy consumption prediction based on double-branch input, and stores computer executable instructions.

Claims

1. A vehicle energy consumption prediction method based on dual-branch input, characterized in that, The method includes: Acquire real-time driving energy consumption data of the target vehicle and perform multimodal feature extraction on the driving energy consumption data to obtain the normalized time series sequence of the driving energy consumption data; Perform multi-level discrete wavelet transform on the normalized time series to determine the standard multi-channel input tensor; By using a pre-defined convolutional neural network, local features of the standard multi-channel input tensor are extracted to obtain a local spatiotemporal feature sequence; The hidden state sequence is determined by calculating and splicing the hidden states at each time step of the local spatiotemporal feature sequence through a long short-term memory network. Calculate the attention weights corresponding to each alignment score in the hidden state sequence to obtain the context vector corresponding to the hidden state sequence; Based on the normalized time series, the macroscopic operating condition feature vector of the target vehicle is determined, and based on the macroscopic operating condition feature vector and the context vector, the energy consumption feature vector of the target vehicle under the driving conditions is determined by vector dimension alignment and Transformer encoding. The energy consumption feature vector is subjected to feature aggregation and energy consumption prediction mapping to obtain the energy consumption prediction value of the target vehicle.

2. The vehicle energy consumption prediction method based on dual-branch input according to claim 1, characterized in that, The driving energy consumption data includes: vehicle speed time series corresponding to each independent operating condition, and total energy consumption corresponding to each independent operating condition; Multimodal feature extraction is performed on the driving energy consumption data to obtain a normalized time-series sequence of the driving energy consumption data, specifically including: Filter the global maximum and global minimum values ​​corresponding to the vehicle speed time series, and calculate the normalized value corresponding to each vehicle speed time element in the vehicle speed time series based on the global maximum and global minimum values ​​to obtain the normalized time series input; Calculate the vehicle driving feature vector corresponding to the driving energy consumption data, and integrate the vehicle driving feature vector into a macroscopic operating condition feature vector; wherein, the vehicle driving feature vector includes: average speed, driving distance, and speed standard deviation; Construct sample labels corresponding to the total energy consumption, and integrate the normalized time series input, macroscopic operating condition feature vector and the sample labels to obtain the normalized time series sequence.

3. The vehicle energy consumption prediction method based on dual-branch input according to claim 1, characterized in that, Performing multi-level discrete wavelet transform on the normalized time series to determine the standard multi-channel input tensor specifically includes: The normalized time series is set as the initial approximation coefficient, and the first approximation coefficient corresponding to the initial approximation coefficient is calculated by low-pass filtering and downsampling. The first detail coefficients corresponding to the initial approximation coefficients are calculated through high-pass filtering and downsampling. Iteratively calculate the second approximation coefficient and the second detail coefficient corresponding to the first approximation coefficient until a preset number of decomposition levels is reached, in order to determine the third detail coefficient; Based on the first detail coefficient, the second detail coefficient, and the third detail coefficient, a set of wavelet coefficients is obtained; The wavelet coefficient set is reconstructed by filtering to determine the reconstructed frequency band components; wherein the reconstructed frequency band components include: reconstructed low-frequency components, reconstructed mid-frequency components, and reconstructed high-frequency components. The normalized time sequence and the reconstructed frequency band components are concatenated to determine the standard multichannel input tensor.

4. The vehicle energy consumption prediction method based on dual-branch input according to claim 1, characterized in that, By using a pre-defined convolutional neural network, local features of the standard multi-channel input tensor are extracted to obtain a local spatiotemporal feature sequence, specifically including: The feature sequence of the standard multi-channel input tensor corresponding to the convolution kernel filter is calculated using a preset convolution kernel filter. Based on the feature sequence, a local original feature map corresponding to the standard multi-channel input tensor is obtained through sliding window processing, and the original feature map is normalized to obtain a normalized feature map. Based on the normalized feature map, an activation feature map is obtained through nonlinear transformation; Calculate the pooling value of each sample in the activation feature map at the corresponding time step to obtain the local spatiotemporal feature sequence.

5. The vehicle energy consumption prediction method based on dual-branch input according to claim 1, characterized in that, By using a Long Short-Term Memory (LSTM) network, the hidden states at each time step of the local spatiotemporal feature sequence are calculated and concatenated to determine the hidden state sequence, specifically including: Define the input gate and candidate cell state corresponding to the local spatiotemporal feature sequence; Based on the input gate and candidate cell states, update the cell states corresponding to the local spatiotemporal feature sequences to determine the current cell state; Calculate the output portion of the current cell state, and perform tanh function normalization on the output portion to determine the hidden state corresponding to the local spatiotemporal feature sequence; Obtain the hidden states corresponding to each time step of the local spatiotemporal feature sequence, and concatenate the hidden states to determine the hidden state sequence.

6. The vehicle energy consumption prediction method based on dual-branch input according to claim 1, characterized in that, Calculate the attention weights corresponding to each alignment score in the hidden state sequence to obtain the context vector corresponding to the hidden state sequence, specifically including: Calculate the alignment score corresponding to each time step in the hidden state sequence, and normalize the alignment score using the Softmax function to obtain the score matrix corresponding to all time steps; Based on the score matrix, calculate the attention weights of the hidden state sequence; The hidden state sequence is weighted and summed according to the attention weights to obtain the context vector.

7. The vehicle energy consumption prediction method based on dual-branch input according to claim 1, characterized in that, Based on the macroscopic operating condition feature vector and the context vector, the energy consumption feature vector of the target vehicle under the operating condition is determined through vector dimension alignment and Transformer encoding, specifically including: Using tokens of the same dimension, the sequence format of the macroscopic operating condition feature vector and the context vector is unified to determine the standard encoding vector; wherein, the standard encoding vector includes: standardized macroscopic operating condition feature vector and standardized context vector; The standard encoding vector is transposed in multiple dimensions to obtain tensor split items, and the attention score matrix corresponding to the tensor split items is calculated by scaling dot product attention. Based on the attention score matrix, the recovery tensor corresponding to the standard encoding vector is calculated in parallel, and the recovery tensor is integrated to determine the multi-head attention output; The multi-head attention output is processed by a feedforward neural network to determine the energy consumption feature vector of the target vehicle under the driving conditions.

8. The vehicle energy consumption prediction method based on dual-branch input according to claim 1, characterized in that, The energy consumption feature vector is subjected to feature aggregation and energy consumption prediction mapping to obtain the predicted energy consumption value of the target vehicle, specifically including: The energy consumption feature vector is subjected to feature aggregation to obtain a fused feature vector; The fused feature vector is linearly mapped to the corresponding energy consumption prediction data to obtain the energy consumption prediction value of the target vehicle.

9. A vehicle energy consumption prediction device based on dual-branch input, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to implement a vehicle energy consumption prediction method based on dual-branch input as described in claims 1-8.

10. A non-volatile computer storage medium for predicting vehicle energy consumption based on dual-branch input, storing computer-executable instructions, characterized in that, When the computer is executed, it implements a vehicle energy consumption prediction method based on dual-branch input as described in claims 1-8.

Images