An energy consumption prediction method based on multi-source physical characteristics and multi-scale recursive convolution

Abstract

The present application relates to new energy automobile energy consumption management and artificial intelligence prediction technical field, specifically relates to a kind of energy consumption prediction method based on multi-source physical characteristics and multi-scale recursive convolution, comprising: step S1: multi-source physical characteristics acquisition and joint embedding;Step S2: multi-scale down-sampling and recursive trend injection;Step S3: local slice convolution and heterogeneous feature fusion;Step S4: double-path deep feature fitting and residual compensation;Step S5: result reconstruction and energy consumption prediction output.The present application has the technical effect that by constructing multidimensional collaborative time feature extraction system, the precision and robustness of energy efficiency prediction under complex physical working conditions are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of energy consumption management and artificial intelligence prediction technology for new energy vehicles, specifically to an energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution, which is particularly suitable for high-precision energy consumption modeling using road network topology under sparse trajectory data conditions. Background Technology

[0002] Traditional energy consumption prediction methods are mainly divided into white-box methods based on physical models and black-box methods based on data-driven approaches. Physical model methods rely on the vehicle's longitudinal dynamics equations and require precise acquisition of parameters such as rolling resistance coefficient, air resistance coefficient, transmission efficiency, and road gradient. However, in actual road driving, these physical parameters are greatly affected by the dynamic changes in ambient temperature, road surface conditions, and vehicle aging, making them difficult to obtain accurately in real time. This results in poor generalization ability of the model under complex and variable real-world conditions. Data-driven methods, on the other hand, utilize massive amounts of historical driving data (such as speed, acceleration, current, voltage, and SOC) to uncover energy consumption patterns through machine learning or deep learning algorithms. Because they do not rely on complex physical parameters and have strong nonlinear fitting capabilities, they have gradually replaced physical models as the mainstream research method in recent years.

[0003] First, energy consumption data exhibits extremely complex non-stationarity and multi-scale fluctuation characteristics. Taking vehicle energy consumption or industrial electricity consumption as examples, it is affected by both macroscopic time-cycle patterns and instantaneous abrupt changes in operating conditions (such as equipment start-up and shutdown, rapid acceleration). Traditional energy efficiency prediction models (such as ARIMA or simple linear regression) struggle to capture such highly nonlinear dynamic changes. Second, existing deep learning models have shortcomings in mining long-term dependencies and local feature balance. While recurrent neural networks (RNNs) and their variants (such as LSTM and GRU) can handle sequential data, they are prone to gradient vanishing problems when dealing with extremely long sequences, and their sequential computation is inefficient. Although traditional temporal convolutional networks (TCNs) can obtain a larger receptive field through dilated convolutions, their feature extraction scale is singular, often capturing global evolutionary trends while ignoring subtle energy consumption fluctuations within local time slices. Furthermore, to further refine feature representation, existing research has generally introduced attention mechanisms or embedding techniques to fuse time features such as hours and dates with the original energy consumption signal in a high-dimensional representation, attempting to improve the model's predictive performance in complex environments. Additionally, existing multi-scale models typically involve simple parallel feature concatenation, lacking deep fusion and recursive feedback mechanisms between features of different granularities. This results in "coarse-grained" trend information failing to effectively guide "fine-grained" detail capture. Finally, noise and valid abrupt changes in energy consumption signals are often difficult to distinguish. In complex physical environments, energy consumption data often exhibits random fluctuations. Existing single-path convolutional architectures are prone to "over-smoothing" of features when extracting high-order nonlinear features, leading to decreased model sensitivity in predicting energy consumption peaks or abnormal loss points.

[0004] The existing technology primarily employs a time series forecasting architecture based on multi-resolution analysis. Its core reference is the Sample Convolution and Interaction Network (SCINet) proposed by Liu et al. in "SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction" and its derivative applications in energy load forecasting. While existing technology one achieves multi-resolution analysis through multi-level downsampling and sub-sequence interaction mechanisms, it still exhibits the following significant shortcomings in practical applications. First, its recursive feature evolution logic mainly focuses on horizontal interactions between sub-sequences at the same level, lacking a vertically guided injection mechanism from macroscopic trends to microscopic fluctuation details. This results in the model failing to provide effective global trend constraints for microscopic instantaneous energy consumption pulses when processing energy consumption data, as the macroscopic energy evolution background cannot provide effective global trend constraints. Second, this technology often uses simple linear accumulation or fixed-weight concatenation methods for feature fusion at different granularities, failing to consider the differences in the demand for local features and global trends during stable and abrupt energy consumption periods. It lacks a dynamic gating fusion method that can adaptively adjust the proportion of heterogeneous features based on real-time input states. Furthermore, the convolutional units relied upon by this scheme are mostly standard single-path architectures, which are prone to feature oversmoothing when mining high-order nonlinear patterns in energy consumption data. Due to the lack of path-level residual compensation for the linear components of the original signal, the model often exhibits significant loss of detail or peak fitting lag when facing complex energy loss abrupt change signals.

[0005] Existing technology two primarily employs a temporal convolutional network architecture based on dilated convolutions for energy consumption sequence modeling. This approach references the improved TCN prediction model proposed by Hussain et al. in "Temporal Convolutional Networks Applied to Energy-Related Time Series Forecasting" and its application in energy systems. While existing technology two effectively expands the temporal receptive field through dilated convolutions, it still has significant shortcomings when processing highly non-stationary energy loss signals containing transient abrupt changes. First, this architecture uses a single-path convolutional feature extraction method. In the process of stacking multiple residual blocks to mine higher-order nonlinear laws, it is prone to "oversmoothing" of the feature signal. Due to the lack of a path-level explicit compensation mechanism for the linear components of the original energy consumption signal, the model's response sensitivity to extreme energy consumption fluctuations and abnormal loss signals is limited. Second, the feature fusion dimension of the standard TCN architecture is relatively simple, mainly relying on the stacking of convolutions in the temporal dimension. It lacks dynamic gating methods that can adaptively adjust weights for local fine slices and global evolution trends, resulting in insufficient ability to coordinate multi-granularity features under complex and variable operating conditions. Furthermore, the expansion of its receptive field is highly dependent on the exponential growth of the expansion factor. It lacks explicit modeling methods for the "coarse-to-fine" recursive evolution logic in the energy consumption sequence, making it difficult to accurately capture the guiding relationship between macro trends and micro details.

[0006] Existing technology three mainly adopts a long-cycle energy consumption prediction architecture based on a probabilistic sparse self-attention mechanism. This technical solution references the Informer model proposed by Zhou et al. and its typical applications in electricity and energy consumption prediction. Although existing technology three has significant advantages in mining global long-range dependencies, it still reveals obvious shortcomings in dedicated energy loss prediction tasks. First, the feature extraction process of this architecture highly depends on the calculation of global attention weights, lacking a direct modeling method for the explicit recursive evolution logic of "coarse to fine" in the energy consumption sequence. This results in the macro-energy evolution background being unable to effectively guide the prediction of micro-fluctuations through parameter injection, leading to low vertical information transmission efficiency. Second, this scheme suffers from insufficient perception accuracy when processing local energy consumption pulse signals and lacks an adaptive gating fusion mechanism that can dynamically adjust the proportion of local detailed features and global trends according to operating conditions, easily causing significant prediction drift at the peaks or troughs of energy consumption fluctuations. Finally, the model's complex self-attention interaction mechanism lacks a path-level explicit compensation design for the linear components of the original physical signal when capturing high-order nonlinear laws of energy consumption data. This results in limited robustness and detail fitting ability of the model when facing energy consumption scenarios with variable operating conditions or poor data quality. Summary of the Invention

[0007] This invention aims to address the shortcomings in energy loss prediction tasks, such as a lack of recursive evolution logic, a single heterogeneous feature fusion method, and excessive smoothing of deep features, by providing a multi-dimensional collaborative temporal feature extraction scheme.

[0008] First, this invention aims to address the problem of limited prediction accuracy in existing technologies when dealing with multi-source dynamic and electrothermal physical variables such as velocity, acceleration, temperature difference, and voltage difference, due to the lack of direct guidance from macroscopic trends to microscopic fluctuations. By constructing a "coarse-to-fine" recursive injection mechanism, the macroscopic energy efficiency background is transformed into vertical physical constraints, ensuring that the model can accurately simulate the explicit recursive evolution logic of these complex physical characteristics at different time scales.

[0009] Secondly, this invention aims to solve the technical problem that existing technologies are too simplistic in their approach to integrating local details such as speed changes and voltage fluctuations with global trends, and cannot adaptively adjust weights. By introducing an adaptive fusion method based on a gating mechanism, the contribution ratio of local high-frequency features and global background information can be dynamically allocated according to real-time dynamic conditions, thereby improving the model's predictive stability and accuracy at moments of sudden energy consumption changes.

[0010] Furthermore, this invention aims to address the problem that existing technologies tend to lose the linear characteristics of the original physical signal when mining high-order nonlinear laws such as speed and temperature difference. By designing a parallel single-path and dual-path residual compensation architecture, while deeply fitting complex energy consumption laws, the linear components of the original physical signal are explicitly preserved and compensated, thereby significantly improving the model's response sensitivity and detail capture capability to energy efficiency peaks and loss points caused by temperature or voltage differences.

[0011] Ultimately, this invention aims to construct an efficient time feature extraction system by integrating multi-scale recursion, gated heterogeneous fusion, and path compensation technologies, in order to comprehensively address the complex dynamic characteristics of energy consumption sequences, which include highly non-stationary, multi-period overlap, and long-range dependence, including dynamic and electrothermal multidimensional variables.

[0012] The specific technical solution is as follows:

[0013] An energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution includes the following steps:

[0014] Step S1: Multi-source physical feature acquisition and joint embedding. The system acquires a multi-dimensional physical sequence in real time, including historical energy consumption E, velocity v, acceleration a, temperature difference ΔT, and voltage difference ΔV, through sensors.

[0015] Step S2: Multi-scale downsampling and recursive trend injection. The embedded high-dimensional feature map is fed into the MDM module, and a multi-resolution feature layer from coarse to fine is constructed through average pooling operations of different lengths. Recursive addition is used to achieve the vertical guidance of macro trends on micro fluctuations.

[0016] Step S3: Local slice convolution and heterogeneous feature fusion. Local temporal slice features and global multi-scale features are extracted in parallel, and their weights are dynamically adjusted through an adaptive gating mechanism.

[0017] Step S4: Dual-path deep feature fitting and residual compensation. Linear features and higher-order nonlinear patterns are extracted using single and dual parallel paths respectively, preventing over-smoothing of features.

[0018] Step S5: Result Reconstruction and Energy Consumption Prediction Output. The processed feature map is transposed and flattened, and the predicted energy consumption value within the target time step is output through linear projection.

[0019] First, we define the following:

[0020] Definition: Input feature space definition

[0021] The defined original input feature matrix is Where N represents the number of observation nodes, and D represents the feature dimension. The feature dimension D includes the original energy consumption sequence E, real-time velocity v, real-time acceleration a, temperature difference ΔT, and voltage difference ΔV, which are multi-dimensional physical variables.

[0022] The purpose of energy consumption prediction is to learn a mapping function based on historical spatiotemporal feature sequences X. The physical states of the past T steps are mapped to the predicted energy consumption sequence of the future T' steps, as shown below:

[0023]

[0024] S1. Multi-source heterogeneous feature acquisition and joint embedding: Original energy consumption sequences and associated physical feature data are acquired separately. The acquired feature data are mapped to a high-dimensional space through a linear projection layer. The calculation process is as follows:

[0025]

[0026] In the formula, The weights are learnable linear mappings.

[0027] S2, Multi-scale downsampling and "coarse-to-fine" recursive trend injection;

[0028] A recursive structure called an IS tree is constructed. This structure effectively captures short-term and long-term temporal correlations at different time resolutions through multi-level downsampling. First, the input sequence is downsampled at k levels using the average pooling operator, generating downsampling rates of... Feature sequence set Subsequently, a "recursive injection" mechanism is adopted, starting from the lowest resolution layer S1, and the following operations are performed level by level:

[0029]

[0030] In the formula, W i Here, σ is the mapping parameter, and σ is the GELU activation function. For each layer, it is a bias term that allows the activation function to output a non-zero value when the input is zero, enabling the model to learn relevant features better.

[0031] S3. Adaptive Gated Fusion of Local Slice Convolution and Heterogeneous Features: An adaptive gated fusion module is introduced, utilizing Local Slice Convolution (TPC) to extract short-term slice features F. local The output of the MDM module is used as the global feature F. global The gate weight g is generated by concatenating the two vectors:

[0032]

[0033] In the formula, σ represents the Sigmoid activation function. This indicates a splicing operation. The symbol represents the characteristics after fusion, and ⊙ represents the Hadamard product.

[0034] S4. Dual-path deep feature fitting and linear residual compensation; the fused features are fed into the temporal convolution module. Parallel single and dual-path designs are implemented.

[0035] Dual convolution path: Extracts high-order nonlinear mappings through two consecutive layers of dilated convolutions.

[0036]

[0037] Single convolution path: Uses a single layer of convolution to directly preserve the linear components of the original physical signal.

[0038]

[0039] Residual merging: The outputs of the two paths are concatenated using 1x1 convolution and residual compensation is performed to obtain the final depth temporal feature Y. TCN :

[0040]

[0041] In the formula, This represents the convolution operation. This indicates a splicing operation. and These represent two different activation functions. The residual compensation is used to stabilize the training process and maintain the integrity of information transmission.

[0042] S5. Result Reconstruction and Multi-Step Energy Consumption Prediction Output; Y TCN After dimension transposition and flattening, the final predicted value Y is output through a linear mapping layer:

[0043]

[0044] The mean absolute error (MAE) is used as the loss function for end-to-end training, and its calculation formula is as follows:

[0045] .

[0046] The technical advantage of this invention is that by constructing a multi-dimensional collaborative time feature extraction system, the accuracy and robustness of energy efficiency prediction under complex physical conditions are significantly improved.

[0047] First, the multi-scale recursive trend injection module was used to realize the vertical physical guidance of macro trends on micro fluctuations. Through the step-by-step information injection mechanism from coarse to fine, the logical discontinuity between macro energy consumption background and micro pulse details was effectively solved.

[0048] Secondly, by integrating local slice convolution and adaptive gating fusion mechanism, the model can dynamically adjust the weight ratio of local and global features according to the physical characteristics such as speed, acceleration, temperature difference and voltage difference in real time input, which greatly enhances the instantaneous response capability and adaptive perception sensitivity of the model at the moment of rapid acceleration or sudden change of working conditions.

[0049] Furthermore, the dual-path residual compensation architecture designed in this invention effectively overcomes the problems of feature oversmoothing and physical signal loss caused by convolution stacking in traditional deep networks by extracting deep nonlinear laws and shallow linear features in parallel and implementing path-level residual compensation, thus ensuring accurate capture of energy consumption peaks and loss points caused by abnormal electrothermal environment.

[0050] Ultimately, this invention establishes a deep characteristic characterization between physical state and energy conversion efficiency by deeply coupling dynamics and electrothermal multidimensional variables, which has extremely high practical value and technical barriers in the fields of new energy vehicles and industrial energy storage prediction. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method of the present invention;

[0052] Figure 2This is a schematic diagram of the MDM module proposed in this invention;

[0053] Figure 3 This is a schematic diagram of the heterogeneous feature fusion module proposed in this invention;

[0054] Figure 4 This is a schematic diagram of the dual-path convolution module proposed in this invention;

[0055] Figure 5 A schematic diagram of the prediction paradigm adopted for the experiment;

[0056] Figure 6 This is a schematic diagram of the experimental data processing method;

[0057] Figure 7 A schematic diagram of the predicted fluctuations in the experimental results.

[0058] Figure 8 A schematic diagram of the predicted fluctuations in the experimental results.

[0059] Figure 9 A schematic diagram of the predicted fluctuations in the experimental results.

[0060] Figure 10 This is a schematic diagram comparing the fitted curves of the energy consumption prediction from the experimental results. Detailed Implementation

[0061] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments are not intended to limit the scope of the present invention.

[0062] The entire technical solution in this embodiment is implemented by five core modules: multi-source heterogeneous feature embedding, multi-scale recursive trend injection module (MDM), local and global feature heterogeneous fusion module (HFF), dual-path residual compensation temporal convolution module (TCN), and high-dimensional mapping output module.

[0063] This specific embodiment provides an energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution. The following definitions are first provided:

[0064] Definition: Input feature space definition

[0065] The original input feature matrix defined in this specific embodiment is: Here, N represents the number of observation nodes, such as the sensor locations of different vehicles or battery modules, and D represents the feature dimension. Unlike traditional studies that only consider historical energy consumption, the feature dimension D in this embodiment includes multi-dimensional physical variables such as the original energy consumption sequence E, real-time velocity v, real-time acceleration a, temperature difference ΔT, and voltage difference ΔV. This definition aims to establish a nonlinear semantic relationship between dynamic state and energy loss through the coupling of multi-source physical signals.

[0066] The purpose of energy consumption prediction is to learn a mapping function based on historical spatiotemporal feature sequences X. The physical states of the past T steps are mapped to the predicted energy consumption sequence of the future T' steps, as shown below:

[0067]

[0068] Based on the above definition, the specific implementation steps provided by this invention are as follows, and the flowchart is shown below. Figure 1 As shown:

[0069] S1. Multi-source heterogeneous feature acquisition and joint embedding: Original energy consumption sequences and associated physical feature data are acquired separately. The acquired feature data are mapped to a high-dimensional space through a linear projection layer. The calculation process is as follows:

[0070]

[0071] In the formula, These are learnable linear mapping weights. This step aligns heterogeneous physical signals, such as velocity units and voltage units, to the same high-dimensional semantic space, laying the foundation for subsequent feature mining.

[0072] S2, Multi-scale downsampling and "coarse-to-fine" recursive trend injection; This embodiment constructs a tree called IS (corresponding to the MDM module, the specific process is as follows) Figure 2 The recursive structure (as shown) effectively captures short-term and long-term temporal correlations at different time resolutions through multi-level downsampling. First, the input sequence is downsampled at k levels using the average pooling operator, generating downsampling rates of... Feature sequence set Subsequently, employing the core "recursive injection" mechanism of this invention, starting from the lowest resolution layer S1, the following operations are performed level by level:

[0073]

[0074] In the formula, W i Here, σ is the mapping parameter, and σ is the GELU activation function. For each layer, it is a bias term that allows the activation function to output a non-zero value when the input is zero, enabling the model to learn relevant features better.

[0075] This process ensures that low-frequency physical laws, such as long-term velocity evolution, can be predicted by guiding high-frequency energy consumption pulses step by step through the longitudinal path, effectively solving the problem of macro-micro logic discontinuity.

[0076] S3. Adaptive Gating Fusion of Local Slice Convolution and Heterogeneous Features: To balance instantaneous changes in energy consumption conditions, such as rapid acceleration or voltage drops, with the global evolution background, this embodiment introduces an adaptive gating fusion module, such as... Figure 3 As shown, short-time slice features F are extracted using Local Slice Convolution (TPC). local This feature is extremely sensitive to minute fluctuations in external factors such as speed and voltage difference. The output of the MDM module is used as the global feature F. global The gate weight g is generated by concatenating the two vectors:

[0077]

[0078] In the formula, σ represents the Sigmoid activation function. This indicates a splicing operation. The symbol represents the characteristics after fusion, and ⊙ represents the Hadamard layer.

[0079] This gating mechanism can automatically adjust the g value according to real-time physical conditions, thereby giving higher weight to local physical features.

[0080] S4. Dual-path deep feature fitting and linear residual compensation; the fused features are then fed into the temporal convolution module. To address the signal smoothing problem that easily occurs when deep networks extract higher-order laws such as acceleration, this embodiment designs parallel single and dual paths, such as... Figure 4 As shown:

[0081] Dual convolution path: Extracts high-order nonlinear mappings through two consecutive layers of dilated convolutions.

[0082]

[0083] Single convolution path: Uses a single layer of convolution to directly preserve the linear components of the original physical signal.

[0084]

[0085] Residual merging: The outputs of the two paths are concatenated using 1x1 convolution and residual compensation is performed to obtain the final depth temporal feature Y. TCN :

[0086]

[0087] In the formula, This represents the convolution operation. This indicates a splicing operation. and These represent two different activation functions. The residual compensation is used to stabilize the training process and maintain the integrity of information transmission.

[0088] This scheme not only deeply fits the nonlinear physical laws, but also explicitly compensates for signal loss, thus improving the accuracy of fitting the peak energy consumption.

[0089] S5. Result Reconstruction and Multi-Step Energy Consumption Prediction Output; Y TCN After dimension transposition and flattening, the final predicted value Y is output through a linear mapping layer:

[0090]

[0091] This embodiment uses Mean Absolute Error (MAE) as the loss function for end-to-end training. Since energy efficiency data is less affected by outliers, MAE provides a more robust linear response and interpretability. Its calculation formula is as follows:

[0092]

[0093] In the comparative experiments, this invention mainly selected three representative algorithmic paths: the first category is statistical models represented by Multiple Linear Regression (MLR) and ensemble learning algorithms represented by XGBoost and LightGBM. These models have extremely high computational efficiency, but their accuracy is limited when processing nonlinear time-series data with complex dynamic laws. The second category is recurrent neural networks represented by LSTM and GRU. They are good at capturing time dependencies and their prediction performance is better than traditional models, but they still have a certain lag in perceiving extreme abrupt changes. The third category is attention mechanism models represented by Transformer. Although they have strong global modeling capabilities, they are prone to feature oversmoothing in energy consumption prediction, resulting in insufficient peak fitting. The model of this invention combines the advantages of the above-mentioned approaches. Through spatiotemporal fusion and recursive injection mechanisms, it achieves optimal prediction accuracy and robustness in both road segment-level and trip-level tasks.

[0094] In terms of data processing and modeling paradigms, this invention abandons the traditional coarse-grained "journey-level" direct prediction method and instead adopts a more physically refined "segment accumulation" prediction paradigm. The differences in prediction process details are as follows: Figure 5 and Figure 6As shown in the diagram. Specifically, this method first dynamically divides a complete long-distance original travel journey into a series of subdivided continuous road segment units based on the topology of the real traffic network. Then, it uses the model of this invention to perform accurate time-series modeling and local energy consumption prediction of multi-source features (such as speed and physical fluctuations) within each road segment unit. Finally, by recursively accumulating and aggregating the prediction results of all micro-road segments in the time domain, a high-precision estimate of the total energy consumption of the entire complete journey is achieved. This "divide and conquer, model segment by segment, and accumulate globally" approach not only effectively captures instantaneous changes in operating conditions between road segments but also significantly improves the model's generalization ability and prediction robustness when facing nonlinear, non-stationary long-distance travel data.

[0095] This invention provides an energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution, which has significant advantages over existing technologies. Comparative experiments on road segment-level and trip-level datasets, as detailed in Tables 1 and 2, demonstrate that this invention exhibits significant advantages in prediction accuracy, robustness, and feature capture capability. Specific beneficial effects are summarized below:

[0096] First, it significantly improves the absolute accuracy of energy consumption prediction under complex operating conditions, effectively solving the problem of large errors in traditional regression models and ensemble learning models when dealing with nonlinear time-series data. Experimental data shows that the mean square error (MSE) of the model in this invention is reduced by approximately 56.5% compared to the traditional MLR model in road segment-level prediction tasks. Compared to the current mainstream ensemble learning model LightGBM, this invention achieves the best results in all three core indicators: mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE). This proves that by introducing multi-source physical features such as velocity, acceleration, temperature difference, and voltage difference, and combining them with high-dimensional projection embedding, this invention can more accurately characterize the underlying coupling relationship between dynamic state and energy loss than traditional black-box models.

[0097] Second, by employing a recursive trend injection and gating fusion mechanism, the model's ability to perceive long-term temporal dependencies and local abrupt changes is significantly enhanced, outperforming mainstream deep learning models. In comparative experiments, the prediction performance of this invention significantly surpasses advanced deep learning models such as LSTM, GRU, and Transformer. Specifically, the RMSE metric of this invention is reduced by approximately 24.2% compared to LSTM and by approximately 28.4% compared to Transformer. This significant improvement is attributed to the "coarse-to-fine" recursive injection mechanism designed in this invention. It injects macroscopic trends as vertical constraints into the prediction of microscopic fluctuations, overcoming the shortcomings of models such as Transformer in terms of insufficient local perception accuracy when processing highly non-stationary energy consumption signals, and ensuring prediction stability at moments of abrupt energy consumption changes (such as acceleration or temperature changes).

[0098] Table 1: Comparison of Prediction Errors at the Road Segment Level

[0099] Table 2: Comparison of travel-level prediction errors

[0100]

[0101] To further evaluate the stability of the model across different test samples or journeys. Figure 7 , Figure 8 and Figure 9 The chart shows the fluctuations in MAE, MSE, and RMSE for each model across 12 different journeys. The red line (marked with a square) representing the STFGCN model is consistently at the bottom of all three subplots. Compared to the dramatic oscillations observed in LightGBM (cyan line) and XGBoost (green line), the STFGCN error curve is extremely smooth. This indicates that STFGCN maintains stable predictive performance regardless of whether energy consumption is at its off-peak or peak, and is less susceptible to fluctuations in individual sample data. Particularly noteworthy is the MSE metric, where STFGCN is almost a straight line, while other models (such as Transformer and GRU) exhibit higher variance. Therefore, the model presented in this invention demonstrates stronger generalization ability when faced with input data of varying distributions.

[0102] Figure 10The line graph visually demonstrates the performance comparison of the proposed model (STFGCN, thick red line) with the actual values ​​(solid gray line) and three mainstream benchmark models in road segment-level energy consumption prediction, strongly proving the absolute advantage of this model in handling highly non-stationary time series data. In terms of global trend tracking, this model consistently runs smoothly close to the actual values ​​throughout the complete sequence, verifying the effectiveness of the multi-scale recursive module (MDM) in long-range macro-trend guidance. Regarding the characterization of local micro-scales, the two magnified local boxes in the figure clearly show that traditional models exhibit significant response lag and amplitude attenuation when facing high-frequency, complex, sawtooth-like micro-fluctuations. At this point, the unique Local Slice Convolution (TPC) module of this invention plays a crucial role. Through the fine-grained receptive field at the slice level, it independently and sensitively extracts the transient high-frequency features of physical quantities such as velocity and voltage, enabling the model to accurately capture every micro-peak and trough with zero delay. Most importantly, near the 100th road segment, the actual energy consumption showed an extreme jump peak approaching 1.2kWh. Traditional deep time series models completely failed due to the over-smoothing effect of features. However, this model successfully and accurately fitted this huge sudden peak by relying on the high sensitivity of the TPC module to capture minute physical anomalies at the bottom layer and the explicit preservation of the original signal by the TCN dual-path residual architecture. The experimental results perfectly demonstrate the core technical barrier of the multi-dimensional feature collaborative extraction system in overcoming the problem of predicting complex variable working conditions.

[0103] Table 3: Computational Efficiency Analysis

[0104]

[0105] Table 3 presents the comparison results of the average training time and inference test time for each model. The experimental results show that the model of this invention (STFGCN), due to its deep multi-scale recursion and graph convolution architecture, has an average training time of 587.815 seconds, which is higher than that of traditional lightweight models. However, it performs excellently in the inference stage, which is of most concern in practical applications.

[0106] In road segment-level inference tests, STFGCN's average response time is only 0.186 seconds, significantly outperforming mainstream deep learning models GRU (0.326 seconds) and LSTM (0.256 seconds). In trip-level aggregation prediction tasks, its test time is further reduced to 0.104 seconds, far exceeding the response speed of ensemble learning models such as XGBoost in this scenario. This demonstrates that the model of this invention has extremely high computational efficiency during the deployment phase, capable of instantaneously processing complex dynamic characteristics and electrothermal physical variables, fully meeting the technical requirements for real-time, online energy consumption monitoring in new energy vehicles or industrial energy management systems.

[0107] In specific embodiments of the present invention, the following technical means may also be employed.

[0108] (1) Input features and embedding layer substitution

[0109] Addition or subtraction of physical feature dimensions: In addition to the speed, acceleration, temperature difference and voltage difference mentioned in this embodiment, alternative solutions may also include ambient humidity, altitude, slope information or motor speed as auxiliary feature inputs.

[0110] Embedding method alternatives: In the feature projection stage, in addition to using a single-layer linear mapping, a multi-layer perceptron or a one-dimensional convolutional layer can also be used for initial feature extraction; for temporal location encoding, in addition to additive fusion, a method of concatenation followed by dimensionality reduction can also be used for representation.

[0111] (2) Alternative steps for multi-scale recursive modules

[0112] Sampling operator substitution: When constructing multi-resolution feature layers, in addition to average pooling, max pooling, dilated convolution, or convolutional layers with strides can be used to reduce the sequence length.

[0113] Recursive injection logic alternative: When injecting macro trends into micro details, in addition to element-level addition, cascaded post-mapping or gated loop units can also be used for feature updates.

[0114] (3) Alternative structure of heterogeneous feature fusion module

[0115] Local feature extraction alternative: Local slice convolution can be replaced by a local window pattern of multi-head self-attention mechanism to capture local dependencies.

[0116] Fusion gating alternatives: The sigmoid activation function used in the gating mechanism can be replaced with Tanh or Softmax for normalized weight allocation; in addition, a simple weighted average or attention score-based weighted fusion can be used to replace the explicit gating structure.

[0117] (4) Path substitution for temporal convolution modules

[0118] Convolution kernel type replacement: The standard convolution in TCN can be replaced with depthwise separable convolution to reduce computation, or with causal convolution to strictly guarantee temporal causality.

[0119] Residual path structure alternatives: The dual-path architecture can be extended to a multi-path parallel architecture (such as a three-path architecture), corresponding to the extraction of linear, low-order nonlinear and high-order nonlinear features respectively; in addition to additive residuals, dense connection mode can also be used for residual connection.

[0120] (5) Replacement of loss function and computational paradigm

[0121] Loss function alternatives: In addition to the MAE loss function used in this embodiment, it can also be replaced with MSE (mean squared error), Huber Loss (smoothed average absolute error), or loss functions with quantile weights according to business needs to cope with energy consumption data with different distributions.

[0122] Output layer substitution: In addition to using fully connected layer mapping, the final reconstructed output layer can also use a decoder structure or a Gaussian process layer to provide uncertainty estimation.

Claims

1. An energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution, characterized in that, Includes the following steps: Step S1: Multi-source physical feature acquisition and joint embedding; The system acquires a multi-dimensional physical sequence in real time, including historical energy consumption E, velocity v, acceleration a, temperature difference ΔT, and voltage difference ΔV, through sensors; Step S2: Multi-scale downsampling and recursive trend injection; The embedded high-dimensional feature map is fed into the MDM module, and a multi-resolution feature layer from coarse to fine is constructed through average pooling operations of different lengths. Recursive addition is used to realize the vertical guidance of macro trends on micro fluctuations. Step S3: Local slice convolution and heterogeneous feature fusion; extract local temporal slice features and global multi-scale features in parallel, and dynamically adjust their weights through an adaptive gating mechanism; Step S4: Dual-path deep feature fitting and residual compensation; linear preservation features and high-order nonlinear laws are extracted through single and dual parallel paths respectively to prevent features from being over-smoothed; Step S5: Result reconstruction and energy consumption prediction output; The processed feature map is transposed and flattened, and the predicted energy consumption value within the target time step is output through linear projection.

2. The energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution according to claim 1, characterized in that, The defined original input feature matrix is Where N represents the number of observation nodes, D represents the feature dimension, and the feature dimension is... D includes the original energy consumption sequence E, real-time velocity v, real-time acceleration a, temperature difference ΔT, and voltage difference ΔV, which are multidimensional physical variables. The purpose of energy consumption prediction is to learn a mapping function based on historical spatiotemporal feature sequences X. The physical states of the past T steps are mapped to the predicted energy consumption sequence of the future T' steps, as shown below: ； S1. The specific method for multi-source heterogeneous feature acquisition and joint embedding is as follows: Original energy consumption sequences and associated physical feature data are acquired separately; the acquired feature data are mapped to a high-dimensional space through a linear projection layer, and the calculation process is as follows: ； In the formula, These are learnable linear mapping weights.

3. The energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution according to claim 2, characterized in that, S2. Multi-scale downsampling and "coarse-to-fine" recursive trend injection, the specific method is as follows: A recursive structure called an IS tree is constructed. This structure effectively captures short-term and long-term temporal correlations at different time resolutions through multi-level downsampling. First, the input sequence is downsampled at k levels using the average pooling operator, generating downsampling rates of... Feature sequence set Subsequently, a "recursive injection" mechanism is adopted, starting from the lowest resolution layer S1, and the following operations are performed level by level: ； In the formula, W i Here, σ is the mapping parameter, and σ is the GELU activation function. For each layer, it is a bias term that allows the activation function to output a non-zero value when the input is zero, enabling the model to learn relevant features better.

4. The energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution according to claim 3, characterized in that, S3. The method of adaptive gating fusion of local slice convolution and heterogeneous features is as follows: an adaptive gating fusion module is introduced, and short-term slice features F are extracted using local slice convolution (TPC). local The output of the MDM module is used as the global feature F. global The gate weight g is generated by concatenating the two vectors: ； ； In the formula, σ represents the Sigmoid activation function. This indicates a splicing operation. The symbol represents the characteristics after fusion, and ⊙ represents the Hadamard layer.

5. The energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution according to claim 4, characterized in that, S4. The method for dual-path deep feature fitting and linear residual compensation is as follows: The fused features are fed into a temporal convolution module; parallel single and dual paths are designed. Dual convolution path: Extracts high-order nonlinear mappings through two consecutive layers of dilated convolutions. ； Single convolution path: Uses a single layer of convolution to directly preserve the linear components of the original physical signal. ； Residual merging: The outputs of the two paths are concatenated using 1x1 convolution and residual compensation is performed to obtain the final depth temporal feature Y. TCN : ； In the formula, This represents the convolution operation. This indicates a splicing operation. and These represent two different activation functions. The residual compensation is used to stabilize the training process and maintain the integrity of information transmission.

6. The energy consumption prediction method based on multi-source physical features and multi-scale recursive convolution according to claim 5, characterized in that, S5. The method for reconstructing the results and generating multi-step energy consumption prediction output is as follows: Y... TCN After dimension transposition and flattening, the final predicted value Y is output through a linear mapping layer: ； The mean absolute error (MAE) is used as the loss function for end-to-end training, and its calculation formula is as follows: 。

Images