New energy automobile energy consumption prediction system and method based on big data

Abstract

The invention provides a new energy automobile energy consumption prediction system and method based on big data, and the system comprises a federal data alignment module which is configured to output an alignment feature vector; the multi-modal feature extraction module is configured to encode the alignment feature vector and output a fusion feature; the dynamic event sensing module is configured to access a road event data stream and output event characteristics and trigger information; the space-time adaptive prediction module is configured to output an energy consumption prediction result based on the fusion feature and the event feature; the credibility evaluation module is configured to generate an evaluation result for indicating prediction reliability; the extreme working condition generation module is configured to form an expanded training set; and the incremental model training module is configured to perform retraining or fine tuning on the prediction model used by the space-time adaptive prediction module. According to the method, on the premise of meeting privacy and compliance constraints, the accuracy and robustness of energy consumption prediction of the new energy vehicle in an extreme / emergency scene are improved.

Description

Technical Field

[0001] This invention belongs to the field of energy consumption prediction technology, specifically relating to a new energy vehicle energy consumption prediction system based on big data. Background Technology

[0002] Energy consumption prediction for new energy vehicles is a fundamental capability for range estimation, trip planning, thermal management, and energy management. Existing technologies generally fall into two categories: one is based on rule / mechanism models of vehicle dynamics and road attributes, estimating energy consumption in segments using parameters such as vehicle mass, drag coefficient, gradient, and temperature; the other is a data-driven approach, training regression or time-series models in the cloud based on historical vehicle operating data (speed / acceleration time series), road grade, and speed limits, and then accumulating energy consumption segment by segment along the route. Some solutions incorporate historical average vehicle speed or static congestion levels to correct the results, or employ centralized data platforms for unified cleaning, offline training, and periodic data distribution.

[0003] The closest existing technology can be summarized as: a cloud-based energy consumption prediction system that uses vehicle-uploaded operational data and road network attributes to generate a unified model, calculates estimated energy consumption in navigation segments, and makes simple corrections based on historical or static traffic information. Some research has also attempted joint / federated training under privacy constraints to reduce the compliance risks of centralized raw data collection. However, these solutions still suffer from the following common problems in complex real-world scenarios, directly affecting the accuracy, timeliness, and usability of predictions: The conflict between data heterogeneity and privacy compliance: Data across vehicle models and suppliers (such as vehicle bus fields, codes, units, and time bases) lacks a unified semantic mapping, leading to frequent instances of "homophones and homonyms." Simultaneously, the centralized aggregation of raw data faces privacy and compliance constraints, making it difficult to achieve high-quality alignment and aggregation without exposing sensitive information. This limits unified modeling and cross-vehicle generalization.

[0004] Insufficient training data coverage: Extreme or long-tail conditions (such as low temperature, high slope, strong crosswind, violent acceleration / deceleration, rare congestion / accidents) are scarce, and many systems lack a mechanism for "when to expand data and how to supplement it", resulting in insufficient representativeness of the training set, poor robustness of the model in rare scenarios, and failure to predict sudden changes in energy consumption.

[0005] Lack of rapid adaptation to emergencies: Traffic accidents, temporary closures, sudden congestion, severe weather, and other events are characterized by strong real-time and time-varying nature. Traditional solutions often rely on static or low-frequency updated road condition indicators for correction, lacking a low-latency linkage path "from event flow to model," making it difficult to adjust inputs or model states in a timely manner, resulting in prediction lag.

[0006] Insufficient spatiotemporal dependency modeling: Many methods only model "single vehicle time series" or "segmented average features", ignoring the structured relationships and adjacency effects between vehicles, road networks and the environment. This makes it difficult to characterize the propagation of events in the road network and their local / global effects on energy consumption, resulting in large prediction biases under complex traffic topologies.

[0007] Model drift and high maintenance costs: Battery aging, seasonal climate change, road network construction, and changes in people's travel patterns can all cause distribution drift. Existing systems mostly rely on periodic offline retraining and full package distribution, lacking incremental learning and miniaturized update paths oriented towards edge computing power constraints, as well as stable version management and differential synchronization mechanisms, resulting in high communication overhead and slow update effectiveness.

[0008] The system suffers from a lack of predictive credibility and interpretability: most systems only output point estimation results and cannot provide confidence intervals or reliability markers. Furthermore, they lack explanatory information from a counterfactual / comparative perspective, making it difficult for the operations and maintenance side to identify high-risk scenarios in a timely manner and take data expansion or strategic intervention measures. At the same time, there is a lack of an engineering path that links the "assessment-interpretation-expansion-update" closed loop.

[0009] Insufficient edge-cloud collaboration and closed-loop capabilities: Existing technologies often process "data cleaning-modeling-deployment" in a fragmented manner, lacking a closed loop that connects "data expansion (supplementary collection / generation)-incremental training-edge distribution-edge replacement" triggered by "evaluation results"; on the edge side, heterogeneous chips and computing power conditions are often incompatible with a single large cloud model, affecting the implementation effect. Summary of the Invention

[0010] The purpose of this invention is to address the shortcomings of the aforementioned background technology and provide a big data-based energy consumption prediction system for new energy vehicles, which improves accuracy and robustness in extreme / sudden scenarios while meeting privacy and compliance constraints.

[0011] The technical solution adopted in this invention is: a new energy vehicle energy consumption prediction system based on big data, comprising communication devices: The federated data alignment module is configured to perform semantic alignment and local cleaning of multi-vehicle and multi-source sensor data under privacy-controlled conditions, and output an aligned feature vector. The multimodal feature extraction module is configured to encode the time-series signal and the environmental signal in the aligned feature vector respectively and output the fused features; The dynamic event perception module is configured to access the road event data stream and output event features and / or trigger information for model updates; The spatiotemporal adaptive prediction module is configured to receive and load the prediction model issued by the edge-cloud collaborative module, and perform vehicle-road network-environment relationship modeling based on the fused features and the event features, and output energy consumption prediction results. The credibility assessment module is configured to quantify the uncertainty of the energy consumption prediction results and generate an assessment result to indicate the reliability of the prediction. The extreme condition generation module is configured to output data expansion instructions to drive supplementary data acquisition and / or conditional data generation based on the assessment results used to indicate the reliability of prediction and / or the data sparsity scenario determined according to the coverage assessment strategy, and to incorporate the obtained supplementary training data into the training dataset to form an expanded training set. The incremental model training module is configured to retrain or fine-tune the prediction model used by the spatiotemporal adaptive prediction module based on trigger information and the expanded training set, and output an updated model. The edge-cloud collaboration module is configured to receive the updated model and send it to the spatiotemporal adaptive prediction module to complete the replacement and take effect.

[0012] In the above technical solution, the system includes end-side nodes instantiated according to the vehicle and at least one cloud-side node; wherein: The federated data alignment module adopts an edge-cloud collaborative deployment. On the edge side, noise suppression, anomaly removal, and semantic / format standardization are performed on local multi-source raw data to form alignment feature vectors. On the cloud side, under privacy-controlled conditions, the alignment feature vectors and their derived data are aggregated and provided to subsequent modules. The multimodal feature extraction module, spatiotemporal adaptive prediction module, and credibility assessment module are deployed on the edge side; The dynamic event perception module is configured to access road event data streams on the cloud side and / or the terminal side and generate event features and / or trigger information. When generated on the cloud side, the information is provided to the terminal side via a communication link. The extreme working condition generation module adopts an edge-cloud collaborative deployment. The cloud side outputs data expansion instructions, and the edge side's acquisition and execution unit implements supplementary data acquisition. The obtained supplementary training data is incorporated into the training dataset to form an expanded training set. The incremental model training module is primarily executed on the cloud side. The edge-cloud collaboration module manages and distributes the updated model on the cloud side, and sends it to the spatiotemporal adaptive prediction module on the edge side to replace and take effect. At the same time, it collects the edge side parameter difference and / or update information as needed for cloud-side aggregation.

[0013] In the above technical solution, the federated data alignment module includes a heterogeneous data mapping unit and a federated alignment unit; The heterogeneous data mapping unit pre-sets and / or dynamically maintains a cross-vehicle field mapping rule library. The semantic standardization maps the original fields, units and codes of different vehicle models to a standard field set and outputs data in a unified format by calling the rule library. The federal alignment unit performs time base alignment, unit normalization, identifier association and missing data processing based on the unified format data, and generates alignment feature vectors under a privacy-controlled collaborative mechanism. The field mapping rule library is called by the extreme working condition generation module and the multimodal feature extraction module to maintain field consistency in the subsequent data processing chain and to perform incremental updates when a new vehicle model or field is added.

[0014] In the above technical solution, the extreme working condition generation module includes an active learning sampling unit and an adversarial generation unit; The active learning sampling unit determines the target operating condition set and generates acquisition trigger information based on the prediction reliability assessment results and / or based on the coverage assessment strategy and / or event characteristics, so that the acquisition execution unit can acquire supplementary data under the target operating condition conditions. The adversarial generation unit uses the description of the target working condition set as the generation constraint to generate synthetic samples corresponding to the target working condition, and supplements the target working condition samples when the supplementary data is insufficient or unavailable. The outputs of the active learning sampling unit and the adversarial generation unit are merged into supplementary training data after consistency and format standardization processing, and then incorporated into the training dataset to form an expanded training set.

[0015] In the above technical solution, the dynamic event perception module includes a real-time traffic condition perception unit and an event response prediction unit; The real-time traffic perception unit receives road event data in a streaming manner and performs parsing, standardization and spatiotemporal alignment to extract event features for energy consumption prediction. The event response prediction unit assesses the immediate and ongoing impact of road events on energy consumption based on the event characteristics, and generates response information for model adaptation. The response information includes at least one item: input correction amount, state / noise parameter adjustment amount, and / or model update trigger information. The event features of the real-time traffic perception unit serve as the input of the event response prediction unit. The response information output by the event response prediction unit is provided to the spatiotemporal adaptive prediction module to perform instant input correction and / or parameter adaptation during the inference process. On the other hand, it is provided to the incremental model training module for updating timing determination and training data selection, and serves as the target condition description of the extreme condition generation module to collaboratively complete data expansion.

[0016] In the above technical solution, the spatiotemporal adaptive prediction module includes a spatiotemporal graph network unit and an adaptive Kalman filter unit; The spatiotemporal graph network unit is used to construct a spatiotemporal relationship representation based on vehicle, road network and environmental elements, aggregate historical driving information and real-time event features, and output prior estimates and event-sensitive features for energy consumption state estimation. The adaptive Kalman filter unit is used to recursively estimate and correct the energy consumption state when receiving the prior estimate, the event-sensitive feature and the fusion feature, and adaptively adjust the state transition and noise statistics when an event-driven state change is detected to mitigate the impact of state transition on the prediction. The filtered state estimate is used as the energy consumption prediction result. The correction residuals and / or parameter update results generated by the adaptive Kalman filter unit are used to provide feedback for adjusting the relationship modeling and aggregation strategy of the spatiotemporal graph network unit.

[0017] In the above technical solution, the multimodal feature extraction module includes a temporal embedding unit and an environment coding unit; The time-series embedding unit is used to perform representation learning on time-series signals related to vehicle operation and extract long-term and short-term features to characterize driving behavior. The environment coding unit is used to learn the representation of environmental and road condition parameters to obtain an environmental representation, and performs time alignment, unit / dimension normalization and feature fusion on the long-term and short-term features and the environmental representation to output fused features; The environment coding unit can perform conditional updates based on the key time period indications provided by the time-series embedding unit to enhance the characterization capability of abnormal or extreme working conditions.

[0018] In the above technical solution, the incremental model training module includes a drift detection unit and an online distillation unit; The drift detection unit is used to statistically identify data distribution and / or concept drift based on runtime data and prediction bias, and to generate trigger information for model updates. The online distillation unit is used to perform knowledge transfer on the target prediction model on the expanded training set and output an updated model after receiving the trigger information. in: The target prediction model refers to the candidate version of the prediction model that is to be loaded and run on the edge by the spatiotemporal adaptive prediction module; The knowledge transfer is based on the supervision signal provided by at least one reference model, which is a version of the same task model maintained by the system. The online distillation unit can determine the sample selection and update range, as well as the capacity and / or structural constraints of the target prediction model, based on the trigger information. The resulting verification results and / or parameter updates are used to provide feedback for adjusting the triggering strategy of the drift detection unit.

[0019] In the above technical solution, the edge-cloud collaboration module includes a model slicing unit and a differential synchronization unit; The model slicing unit is used to generate and / or select model variants based on the updated model as a baseline, according to the computation, storage and communication constraints and task objectives on the end side, and form variant description information for distribution and replacement. The differential synchronization unit is used to perform differential transmission and aggregation between the cloud side and the edge side based on the variant description information under privacy-controlled conditions, synchronize the derived update information of the model, and complete the verification. The model slicing unit determines the target model variant and distributes it to the differential synchronization unit. After verification, the spatiotemporal adaptive prediction module loads the target model variant and replaces it. Update information generated during edge operation is transmitted back by the differential synchronization unit and aggregated on the cloud side. The aggregation result is used to adjust the model slicing and distribution strategy.

[0020] In the above technical solution, the credibility assessment module includes an uncertainty quantification unit and a counterfactual interpretation unit; The uncertainty quantification unit is used to quantify the uncertainty of the energy consumption prediction results and generate an evaluation result that indicates the reliability of the prediction. The counterfactual explanation unit is used to determine the prediction results that need to be explained based on the evaluation results, construct counterfactual samples and / or comparison scenarios related to the prediction results, and output explanation information; The evaluation results are used to determine whether to initiate data expansion. The explanatory information serves as the target constraint and sample selection basis for the extreme working condition generation module to expand the data. The counterfactual samples are incorporated into the training dataset as a candidate for the supplementary training data. The comparison scenario is used to define the reference conditions and / or boundaries of the difference from the prediction results.

[0021] This invention provides a method for predicting the energy consumption of new energy vehicles based on big data, and the system implementation based on the above technical solution includes: Under privacy-controlled conditions, semantic alignment and local cleaning are performed on multi-vehicle and multi-source sensor data to obtain aligned feature vectors. The time-series signal and environmental signal in the aligned feature vector are encoded separately and the fused features are output. Access road event data streams and generate event features and / or trigger information for model updates; Receive and load the prediction model issued by the edge-cloud collaboration module, perform vehicle-road network-environment relationship modeling based on the fused features and the event features, and output energy consumption prediction results; The uncertainty of the energy consumption prediction results is quantified to generate an evaluation result that indicates the reliability of the prediction. Based on the evaluation results used to indicate the reliability of the prediction and / or the data sparsity scenario determined according to the coverage evaluation strategy, output data expansion instructions to drive the acquisition of supplementary data and / or the generation of conditional data, and incorporate the obtained supplementary training data into the training dataset to form an expanded training set; Based on the triggering information and the expanded training set, the prediction model is retrained or fine-tuned to obtain an updated model. The updated model is distributed to the edge side and the replacement takes effect through the edge-cloud collaboration module.

[0022] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the processor performs the steps of the method described in the above technical solution.

[0023] The present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the program is executed by the processor, the electronic device performs the steps of the method described in the above technical solution.

[0024] The beneficial effects of this invention are as follows: This invention constructs a closed-loop, self-evolving technical path: from aligning multi-source data and extracting fusion features, to completing energy consumption inference on the loaded prediction model, then using uncertainty assessment results to drive data expansion and incremental training, and finally completing model replacement through edge-cloud collaboration. This architecture, under the premise of controlled privacy, achieves a simultaneous balance of accuracy, timeliness, and compliance, significantly reducing the frequency of manual intervention and maintenance complexity.

[0025] Furthermore, this invention achieves a balance between real-time performance and centralized optimization by clearly defining the functional division between the edge and cloud sides: the edge side handles latency-sensitive feature extraction, prediction, and reliability assessment, while the cloud side is responsible for alignment aggregation and model update distribution. Using derived data aggregation instead of raw data backhaul reduces bandwidth and compliance risks, while ensuring clear and controllable input / output links between modules, facilitating engineering implementation and maintenance.

[0026] Furthermore, the federated data alignment mechanism of this invention unifies semantics and format through a cross-vehicle model field mapping rule base, and then completes instance-level alignment at the levels of time reference, unit / dimension, and identifier under privacy control, outputting an aligned feature vector. Its effect is to significantly alleviate training bias caused by vehicle model differences, improve cross-vehicle generalization ability, and support incremental maintenance and reuse of the rule base, accelerating the access of new vehicle models / new fields.

[0027] Furthermore, this invention decomposes the data expansion for extreme operating conditions into "active sampling to supplement real data + conditional generation to supplement scarce samples," and uses the credibility assessment result as a directional trigger. This design can quickly improve the representativeness of training data in rare scenarios such as long tails, low temperatures, high slopes, and sudden congestion, enhance the robustness of the model to sudden changes in energy consumption, and improve the convergence efficiency and stability of retraining by incorporating the expanded training set after consistency standardization.

[0028] Furthermore, this invention establishes a low-latency link from road event flow to model adaptation: real-time traffic condition perception outputs event features, and the event response prediction unit generates input correction amounts, state / noise parameter adjustment amounts, and / or update trigger information. Thus, the model can instantly correct inputs and parameters in sudden scenarios such as accidents and congestion, and trigger updates when needed, significantly shortening the "event occurrence - model response" time.

[0029] Furthermore, this invention combines spatiotemporal graph relationship priors with adaptive Kalman filtering: the former aggregates the structured dependencies between vehicles, road networks, and the environment, outputting prior estimates and event-sensitive features; the latter performs recursive corrections upon receiving priors, event features, and fused features, and adaptively adjusts during state transitions. The result is that it can both perceive the propagation effect of road network topology and suppress prediction jitter caused by event disturbances, resulting in smoother and more reliable filtered state estimates.

[0030] Furthermore, this invention unifies the dimensions and time references of driving behavior and external operating conditions through multimodal fusion of temporal embedding and environmental coding, outputting high-quality fused features. Combined with conditional updates based on key time period indications, the system's ability to represent abnormal / extreme operating conditions is enhanced, providing more discriminative input for subsequent prediction and uncertainty quantification.

[0031] Furthermore, this invention constructs an incremental training mechanism oriented towards edge-side constraints, using drift detection as the update trigger and online distillation as a lightweight path. This mechanism can accurately limit the sample range and update amplitude when the distribution changes, transferring the knowledge of the complete model to the deployable target prediction model, maintaining long-term accuracy while controlling edge-side computing power and update costs.

[0032] Furthermore, this invention addresses the contradiction between "vehicle-specific design and low communication overhead" from an engineering distribution perspective: the model slicing unit generates / selects model variants based on end-side constraints, and the differential synchronization unit performs hot replacement after verification using parameter differentiation, sparse updates, or statistical summaries. Update information during runtime is fed back to optimize variant selection and distribution strategies, enabling adaptive closed-loop scheduling of model distribution, shortening deployment latency, and reducing bandwidth consumption.

[0033] Furthermore, this invention provides "credible and usable" capabilities at the output level: uncertainty quantification provides evaluation results to help identify high-risk predictions; counterfactual interpretation, through counterfactual samples and comparative scenarios, constitutes explanatory information, which not only improves auditability but also indicates the expansion targets and sample selection basis for the extreme condition generation module. Thus, data expansion is more accurate, training resources are more focused, and overall evolution is more efficient.

[0034] Overall, this application takes "privacy-controlled data alignment - multimodal representation - spatiotemporal adaptive prediction - reliable assessment - targeted data expansion - incremental training - edge distribution and hot replacement" as its main thread, forming a closed-loop system that runs through data collection, modeling, assessment, expansion, and deployment. This system generates synergistic benefits in five dimensions: accuracy, timeliness, robustness, compliance, and maintainability. It can systematically alleviate the core pain points of existing technologies in terms of cross-vehicle heterogeneity, insufficient long-tail coverage, event adaptive lag, model drift, and edge-cloud deployment. Attached Figure Description

[0035] Figure 1 The topology diagram of the new energy vehicle energy consumption prediction system based on big data of the present invention is shown; Figure 2 The flowchart of the new energy vehicle energy consumption prediction method based on big data of the present invention is shown. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but these descriptions do not constitute a limitation on the present invention.

[0037] Example 1 like Figure 1 As shown, this invention provides a new energy vehicle energy consumption prediction system based on big data, comprising systems that communicate with each other: The federated data alignment module is configured to perform semantic alignment and local cleaning of multi-vehicle and multi-source sensor data under privacy-controlled conditions, and output an aligned feature vector. The multimodal feature extraction module is configured to encode the time-series signal and the environmental signal in the aligned feature vector respectively and output the fused features; The dynamic event perception module is configured to access the road event data stream and output event features and / or trigger information for model updates; The spatiotemporal adaptive prediction module is configured to receive and load the prediction model issued by the edge-cloud collaborative module, and perform vehicle-road network-environment relationship modeling based on the fused features and the event features, and output energy consumption prediction results. The credibility assessment module is configured to quantify the uncertainty of the energy consumption prediction results and generate an assessment result to indicate the reliability of the prediction. The extreme working condition generation module is configured to supplement training data based on the evaluation results generated by the credibility evaluation module to form an expanded training set; The incremental model training module is configured to retrain or fine-tune the prediction model used by the spatiotemporal adaptive prediction module based on trigger information and the expanded training set, and output an updated model. The edge-cloud collaboration module is configured to receive the updated model and send it to the spatiotemporal adaptive prediction module to complete the replacement and take effect.

[0038] At the deployment level, this system includes end-side nodes instantiated according to vehicles and at least one cloud-side node. Each module operates based on an edge-cloud collaborative approach and interacts with data and control commands through communication links.

[0039] The federated data alignment module is primarily responsible for the standardized mapping and federated alignment of multi-source heterogeneous data. Its inputs include various raw data collected during vehicle operation, such as vehicle CAN bus information, GPS positioning data, and environmental sensor data. The module contains two sub-units: a heterogeneous data mapping unit and a federated learning alignment unit. First, the heterogeneous data mapping unit uses a pre-built knowledge graph to semantically standardize the metadata of different vehicle models and sensors, establishing a cross-vehicle data field mapping rule base to unify the raw data format of different vehicle models. After mapping standardization, the data fields uploaded by each vehicle can correspond to each other.

[0040] Subsequently, the federated learning alignment unit performs distributed data cleaning and feature alignment using a federated learning framework. In an optional implementation, the federated learning alignment unit can perform local data cleaning and alignment based on the PySyft framework, uploading only the aligned feature vectors to meet privacy-controlled collaboration requirements.

[0041] The federated alignment unit further performs time-based alignment, unit / dimension normalization, identifier association, and missing data processing based on the unified format data, forming an aligned feature vector under a privacy-controlled collaborative mechanism. The field mapping rule library is called by the multimodal feature extraction module and the extreme condition generation module to maintain field consistency and is incrementally updated when new vehicle models or fields are added. For example, noise filtering and outlier removal are first performed at each local node, and only the normalized feature vector is extracted and uploaded to the cloud for aggregation. This "local cleaning, centralized alignment" approach solves the problem of inconsistent formats from different data sources and protects the privacy of the original data at each node. The output of this module is unified and cleaned vehicle operation feature data, providing standardized input for the subsequent feature extraction module. The cloud side aggregates the aligned feature vector and its derived data (such as statistical summaries, quality markers, and consistency verification information) under privacy-controlled conditions and uses it as a unified input source for subsequent multimodal feature extraction and model training.

[0042] The multimodal feature extraction module receives standardized time-series data and environmental data output from the federated data alignment module, and extracts a comprehensive feature vector for prediction. The module comprises two sub-units: a time-series embedding unit and an environmental coding unit, which process the time-series information of vehicle operation and environmental influencing factors, respectively.

[0043] First, the temporal embedding unit extracts features from the time-series data from the vehicle. For example, it uses temporal models based on the Transformer architecture, such as Informer Temporal Transformer, to encode the sensor sequences of vehicle speed, acceleration, etc., and to mine the long-term driving patterns and short-term dynamic change features.

[0044] Simultaneously, the environment coding unit encodes and fuses relevant parameters of the vehicle's environment. This unit can use convolutional neural networks (such as one-dimensional ResNet) to extract effective information from multi-dimensional environmental variables such as road slope, ambient temperature, and wind speed, converting these environmental parameters into fixed-length environmental feature vectors. Next, the module fuses the temporal embedding features with the environmental features to form a multimodal joint feature representation. During the fusion process, time alignment and unit / dimension normalization are performed to eliminate differences in sampling frequency and measurement units. This joint feature includes both the vehicle's behavioral patterns over time and external environmental conditions, providing a comprehensive characterization of energy consumption. The environment coding unit can perform conditional updates based on key time period indications provided by the temporal embedding unit to enhance its ability to represent abnormal or extreme operating conditions.

[0045] The output of the multimodal feature extraction module is a high-dimensional sequence of feature vectors, which serves as the input basis for the spatiotemporal adaptive prediction module, providing rich information support for energy consumption prediction. This module is deployed on the edge node, and its output fused features are directly used by the edge spatiotemporal adaptive prediction module.

[0046] The dynamic event perception module focuses on incorporating dynamic event information from outside the vehicle, perceiving sudden road conditions and traffic events, and promptly feeding this information back into the prediction process. The input to the dynamic event perception module primarily comes from real-time traffic data sources, such as accessing the Gaode / Baidu real-time traffic API through the Apache Flink stream processing engine to obtain road event information streams. The module comprises two sub-units: a real-time traffic perception unit and an event response prediction unit. First, the real-time traffic perception unit receives event data in a streaming manner and performs parsing, standardization, and spatiotemporal alignment to extract event features for energy consumption prediction. The real-time traffic perception unit continuously monitors and parses data from the traffic API, including information such as road congestion levels, accident locations, and times. By processing this real-time event data, it extracts key event features that affect vehicle energy consumption (e.g., severe congestion or an accident ahead).

[0047] Then, these event features are passed to the event response prediction unit as additional input or correction signals. Based on the received event features, the event response prediction unit dynamically adjusts the relevant parameters or input vectors of the energy consumption prediction model, updates the prediction in real time to respond to the latest events, and generates response information for model adaptation. This response information includes at least: input correction amounts, state / noise parameter adjustments, and / or model update trigger information. This process can be implemented through an online learning algorithm, that is, by slightly adjusting the weights of the energy consumption model based on newly emerging events during system operation, thereby ensuring the model's sensitivity and response speed to sudden events. Due to the adoption of efficient streaming data processing and online update strategies, this module controls the additional response latency to the millisecond level (e.g., within 200ms), enabling it to quickly reflect the impact of sudden road condition changes on vehicle energy consumption in the prediction results without significantly increasing the computational burden. The dynamic event perception module outputs event-corrected features or updated model parameters, allowing subsequent prediction modules to provide more accurate energy consumption predictions based on the latest road condition information. The event features and / or response information can also serve as the target operating condition description for the extreme operating condition generation module, used to collaboratively complete data expansion.

[0048] The dynamic event perception module can be deployed on the cloud side and / or the edge side: when the road event data stream is accessed and parsed on the cloud side, the generated event features and / or triggering information are provided to the edge side via the communication link; when accessed on the edge side, it can be directly used for edge side prediction and evaluation.

[0049] The spatiotemporal adaptive prediction module is the core of the energy consumption prediction system, used for spatiotemporal correlation prediction of vehicle energy consumption based on multimodal features and real-time event information. Its inputs include the comprehensive feature vector output by the multimodal feature extraction module and the latest event correction information provided by the dynamic event perception module. Internally, the spatiotemporal adaptive prediction module consists of a spatiotemporal graph network unit and an adaptive Kalman filter unit. The correction residuals and / or parameter update results generated by the adaptive Kalman filter unit are used to adjust the relationship modeling and aggregation strategy of the spatiotemporal graph network unit.

[0050] First, the spatiotemporal graph network unit constructs a heterogeneous spatiotemporal relationship model among vehicles, road networks, and the environment to capture the complex dependencies of energy consumption over time and space. Specifically, graph neural network technology (such as the GraphSAGE algorithm) can be used to fuse historical vehicle trajectory data with real-time road conditions and environmental states into a spatiotemporal heterogeneous graph. In this graph network, nodes can represent vehicle states or road nodes, and edges represent vehicle travel paths or interaction relationships. The graph neural network iteratively calculates this heterogeneous graph, aggregating the feature information of neighboring nodes, and finally outputs a predicted value for the current vehicle energy consumption. With the help of the graph network unit, the system can simulate complex situations such as the impact of local road network congestion on vehicle energy consumption and the spatial diffusion of energy consumption patterns of vehicle groups, thereby predicting phenomena such as sudden changes in vehicle energy consumption caused by sudden congestion.

[0051] Next, the adaptive Kalman filter unit performs online correction and smoothing of the initial prediction results output by the graph network. Considering that traffic events may cause sudden changes in the model state, this unit introduces an adaptive estimation mechanism based on the classic Kalman filter: dynamically adjusting the noise covariance matrix in the filtering process using the variational Bayesian method, enabling the filter to automatically increase or decrease its confidence in new observations according to the current operating conditions. In this way, when sudden changes in road conditions cause a sudden increase in prediction error, the adaptive filter unit will promptly increase the weight of new data and quickly correct the prediction bias; while under normal and stable driving conditions, it maintains confidence in the model's prior knowledge to avoid excessive fluctuations.

[0052] After processing by the two units mentioned above, the spatiotemporal adaptive prediction module outputs the final vehicle energy consumption prediction result (which can be an instantaneous energy consumption value or an energy consumption curve over a period of time). These prediction results are then sent to the reliability assessment module for evaluation of the reliability of the results. This module is deployed on the edge node and receives and loads the prediction model and / or model variants issued by the edge-cloud collaboration module as the currently active model for inference.

[0053] The credibility assessment module quantifies the uncertainty of energy consumption prediction results and interprets the results, aiming to evaluate the credibility of the predictions and provide interpretable evidence for the results. The credibility assessment module accepts energy consumption prediction results from the spatiotemporal adaptive prediction module, along with relevant input feature information. The module consists of an uncertainty quantification unit and a counterfactual interpretation unit.

[0054] First, the uncertainty quantification unit provides a confidence assessment of the prediction result at each time step. For example, uncertainty estimation methods such as Monte Carlo Dropout can be introduced into the prediction model to perform multiple random Dropout inferences on the same input, thereby obtaining the distribution characteristics of the prediction results and calculating the confidence interval of the energy consumption prediction. By analyzing the width of the confidence interval, the system can determine the credibility of the current prediction: if the confidence interval is narrow, it indicates that the prediction is relatively stable and reliable; conversely, if the confidence interval is wide, it indicates that the prediction has significant uncertainty and needs to be treated with caution. For energy consumption prediction results marked as low confidence, the module will trigger further interpretation and analysis.

[0055] Subsequently, the counterfactual explanation unit generates corresponding counterfactual samples for these low-confidence scenarios and provides comparative analysis. Counterfactual samples refer to the predicted output of vehicle energy consumption under slightly altered current conditions, helping to understand the sensitivity of the original prediction results to changes in different factors. This unit can use a neural network generative model (such as a neural process model) to generate energy consumption results under several hypothetical scenarios based on the current input. For example, when a prediction shows that a vehicle's energy consumption is abnormally high under specific conditions with low confidence, the counterfactual explanation unit can generate a hypothetical energy consumption result under slightly lower temperature or gentler slope conditions, compare it with the original result, and explain the possible reasons for the increased energy consumption. Through uncertainty quantification and counterfactual analysis, this module outputs a confidence score and explanation report for each energy consumption prediction. On the one hand, the confidence assessment results can indicate to the system or user which predictions require attention or further data collection and verification; on the other hand, this information will be fed back to guide subsequent data expansion and model optimization processes. This module is deployed on edge nodes, and its output assessment results are used for local risk warnings and to trigger / constrain subsequent data expansion and model update processes. The counterfactual explanation unit constructs counterfactual samples and / or comparison scenarios related to the prediction result and outputs explanation information; wherein, the comparison scenarios are used to define reference conditions and / or boundaries that differ from the prediction result. The evaluation result is used to determine whether to initiate data expansion; the explanation information serves as the target constraint and sample selection basis for the extreme condition generation module to perform data expansion, and the counterfactual samples can be incorporated into the training dataset as a type of candidate supplementary training data.

[0056] The extreme condition generation module is used to augment data for scenarios with sparse or low-reliability data by synthesizing or collecting samples under extreme conditions to enhance model training. It is typically triggered by the reliability assessment module and historical data analysis results: when insufficient data samples are found under certain conditions, or when the model's prediction uncertainty is high under certain special circumstances, the system will activate the extreme condition generation module to enrich the data for these "weak" scenarios. The module internally consists of two parts: an active learning sampling unit and an adversarial generation unit.

[0057] First, the active learning sampling unit is responsible for identifying insufficiently covered operating conditions and actively collecting real data. This unit can dynamically filter out areas not included or with very few samples in the current training dataset from all possible operating condition spaces using active learning strategies such as uncertainty sampling. For example, monitoring historical data shows that there is very little data in the combined scenario of extreme low temperatures (such as below -10℃) and high-speed vehicle acceleration, and the prediction model has low confidence under these conditions. In this case, the active learning sampling unit would identify this as an uncovered extreme operating condition.

[0058] Subsequently, commands are issued via vehicle-to-everything (V2X) communication or over-the-air (OTA) updates to trigger eligible vehicles to collect data for that scenario during actual operation. For example, the system can instruct vehicles to simulate rapid acceleration or activate battery heating under safe conditions to obtain energy consumption data under low-temperature, high-speed operating conditions. Through this proactive triggering mechanism, the system can supplement the previously scarce real-world scenario data.

[0059] Next, the adversarial generation unit utilizes generative adversarial networks (GANs) to synthesize simulated extreme operating condition data. Based on the known distribution of real-world data, this unit trains a conditional generative model (such as a conditional Wasserstein GAN), using environmental parameters or operating condition description vectors as inputs to generate physically plausible virtual data samples. For example, using temperature and road gradient as conditions, it generates simulated curves showing the changes in vehicle battery voltage and current over time under extreme conditions. The discriminator of the GAN ensures that the synthesized data is statistically consistent with real data, allowing these synthesized samples to form an expanded training set together with the real-world collected data. Through the combined effect of actively sampled real data and GAN-generated simulated data, the system significantly enhances the sample size and diversity under extreme conditions.

[0060] The extreme condition generation module ultimately outputs an enhanced training dataset, including the original data and newly added extreme scenario samples. This data will be provided to the incremental model training module for updating and training the model, thereby improving the model's prediction accuracy and robustness under rare conditions. The cloud side outputs data expansion instructions and sends them to the edge-side acquisition and execution unit to implement supplementary data acquisition; the acquired supplementary training data and the conditionally generated synthetic samples are processed for consistency and format standardization before being incorporated into the training dataset to form an expanded training set, which is then submitted to the incremental model training module.

[0061] The incremental model training module utilizes a continuously expanding database and newly arriving data streams to perform online updates and adaptive optimization of the energy consumption prediction model, ensuring that the model maintains high accuracy as the environment and operating conditions change. The inputs to the incremental model training module include: first, new samples from the extreme operating condition generation module and real data accumulated during daily operation; second, the latest event impacts continuously provided by the dynamic event perception module; and third, changes in data distribution monitored by the model during actual deployment. The module consists of a drift detection unit and an online distillation unit, each performing its specific function to update the model.

[0062] First, the drift detection unit is responsible for monitoring the statistical characteristics of vehicle energy consumption data in real time, detecting changes in data distribution or vehicle usage patterns. When a significant drift is detected in the distribution of the input data stream (e.g., a significant change in mean or variance identified by drift detection algorithms such as Page-Hinkley), the unit determines that the applicability of the model may have decreased, thus triggering a model update process. At this time, the incremental model training module will retrain or incrementally learn the prediction model using the latest full data (including the original training set and newly added data). Through periodic or real-time small-batch updates, the model can correct prediction biases caused by factors such as environmental changes or vehicle aging. The trigger information generated by the drift detection unit for model updates can be used in parallel or in combination with the trigger information output by the dynamic event perception module.

[0063] Next, the online distillation unit compresses and optimizes the model after the model update to adapt to the resource limitations of the vehicle terminal. Upon receiving the trigger information, the online distillation unit performs knowledge transfer on the target prediction model: the target prediction model refers to the candidate version of the prediction model to be loaded and run on the edge by the spatiotemporal adaptive prediction module; the reference model used for distillation is a system-maintained reference model for the same task (such as a cloud-side aggregated model, historical versions, and / or model integration). Considering that the new model may become complex and large due to the inclusion of a large amount of new data, which is not conducive to real-time operation at the vehicle edge, the online distillation unit uses knowledge distillation technology to transfer the knowledge of the complete model to a simplified sub-model. The verification results and / or parameter updates generated by online distillation are used to provide feedback for adjusting the triggering strategy and sample selection range of the drift detection unit.

[0064] Specifically, a dynamic network structure adjustment method can be adopted: after training a high-precision complete model in the cloud, this model is used as the teacher model to continuously guide the learning of the student model (a smaller model) online, so that it can maintain predictive performance as close as possible to the larger model while significantly reducing the number of parameters and computational load. During the online distillation process, the network width or depth of the student model is dynamically adjusted according to the on-board computing resources, so as to meet the real-time computing requirements of edge devices while taking into account the prediction accuracy.

[0065] Finally, the incremental model training module outputs updated and compressed energy consumption prediction model parameters, providing the entire system with a self-evolving model. After the model update is complete, the new model will be distributed and deployed by the edge-cloud collaboration module. The above incremental model training process is primarily executed on the cloud side, while the edge side can perform lightweight fine-tuning and verification according to resource conditions.

[0066] The edge-cloud collaboration module is responsible for the efficient collaboration and deployment of models between the cloud and vehicles, ensuring timely synchronization of models and data between multiple vehicles and the central server. The module's input includes the latest model parameters generated by the incremental model training module, as well as the model and data status of each vehicle terminal. Internally, the module contains a model slicing unit and a differential synchronization unit. The model slicing unit prunes and distributes the global model according to the computing power of different vehicle terminals.

[0067] Specifically, the system can pre-generate several model variants using techniques such as Neural Architecture Search (NAS), each with different computational complexity and resource consumption. The model slicing unit selects the most suitable model version from these variants based on the computing power and memory size of the chip equipped in the vehicle and distributes it to the vehicle terminal, ensuring that each vehicle runs a model that matches its computing power, thereby maximizing prediction accuracy without sacrificing real-time performance.

[0068] Furthermore, for certain high-computing-power terminals, the model slicing unit also supports an edge-cloud layered deployment scheme. For example, some layers of the deep learning model can be deployed in the cloud and others in the vehicle, further reducing the load on the vehicle and leveraging cloud computing power to improve performance. On the other hand, the differential synchronization unit is used to minimize communication overhead when synchronizing model updates between edge devices and cloud servers. Whenever the local model on the vehicle is updated (e.g., by fine-tuning some weights through online learning) or a new model is obtained through centralized training in the cloud, the differential synchronization unit does not directly transmit the entire model's parameter file. Instead, it calculates the difference in model parameters compared to the previous version and only sends the differing parts. This approach is similar to the application of the Rsync protocol in file synchronization, i.e., only synchronizing the changed parts, significantly reducing network bandwidth consumption. When aggregating model updates from multiple vehicles in the cloud, a differential merging method is also used, integrating only the incremental contributions from each edge device to quickly obtain a unified global model and distribute it.

[0069] Through the combination of model slicing and parameter differential synchronization, the edge-cloud collaboration module achieves efficient distribution and continuous updating of models between the cloud and vehicle terminals: the cloud always stores the latest full model parameters, while each vehicle obtains a model version suitable for itself as needed and maintains regular updates, ensuring that the energy consumption prediction model of the entire fleet is in a state of collaborative evolution. The cloud side performs version management and distribution for updated models, and the differential synchronization unit sends the target model variant to the edge side; the spatiotemporal adaptive prediction module on the edge side loads the target model variant and replaces it after completing the consistency verification; the parameter differential and / or update information generated during the operation of the edge side is returned by the differential synchronization unit and aggregated on the cloud side to optimize the variant selection and distribution strategy.

[0070] In summary, the big data-based new energy vehicle energy consumption prediction system of this invention, through the organic coordination of the aforementioned modules, forms a closed-loop process from data acquisition and preprocessing, feature extraction, real-time prediction to result evaluation and feedback, data expansion, adaptive model updates, and distribution and deployment. This system can integrate multi-source heterogeneous data for high-precision energy consumption prediction while protecting data privacy, and can dynamically adjust prediction results in response to environmental changes and unforeseen events, providing result credibility assessment and interpretability. Furthermore, by continuously enriching extreme scenario data and updating the model online, combined with the synergy of edge computing and cloud computing, this system can continuously improve the generalization performance and robustness of the prediction model, enabling those skilled in the art to fully reproduce and achieve accurate prediction of new energy vehicle energy consumption based on the aforementioned system architecture.

[0071] Example 2 like Figure 2 As shown, this embodiment proposes a method for predicting the energy consumption of new energy vehicles based on the above system. The system deployment includes end-side nodes instantiated by vehicle and at least one cloud-side node: multimodal feature extraction, spatiotemporal adaptive prediction, and reliability assessment are deployed on the end-side; incremental model training is primarily performed on the cloud side; extreme condition generation adopts an end-cloud collaborative deployment; and edge-cloud collaboration is responsible for version management, distribution, and backhaul aggregation. Specifically, it includes the following steps: S1. Semantic alignment and cleaning of multi-source data: Semantic standardization is performed on metadata from multiple sources such as CAN bus, GPS, and environmental sensors. A cross-vehicle field mapping rule library is constructed to unify the original data format of different vehicle models. Noise filtering is performed on local nodes through a distributed data cleaning framework, and only aligned feature vectors are uploaded.

[0072] S1.1 Metadata Standardization: Parse the CAN bus protocol document to extract the physical meaning and units of key fields such as RPM and battery SOC; establish a knowledge graph of sensor metadata and define OWL semantic relationships for fields such as GPS latitude and longitude and ambient temperature; generate cross-vehicle mapping rules in XML format (for example, uniformly mapping the "BMS_SOC" field of each vehicle model to the standard field "battery.state_of_charge").

[0073] S1.2 Distributed Data Cleaning: Kalman filters are deployed at edge nodes to process the raw GPS coordinate data, including velocity constraint verification (removing outliers with speeds greater than 300 km / h) and heading angle smoothing (sliding window size of 5 seconds). Then, PySyft's SecureAggregation protocol is used to aggregate the cleaned feature vectors of each node, automatically removing outliers that exceed the mean ± 3σ range.

[0074] Building upon standardization and local cleaning, federated alignment further performs time-based alignment, unit / dimension normalization, identifier association, and missing data handling to form an aligned feature vector. The field mapping rule base is used by multimodal feature extraction (S5) and extreme condition generation (S2) to maintain field consistency and is incrementally updated when new vehicle models or fields are added. S1.3 Feature Alignment Upload: The cleaned and standardized CAN bus data is converted into the standard Apache Parquet format; when uploaded, only the hashed vehicle VIN code, timestamp, and corresponding feature vector are included to ensure data privacy.

[0075] S2. Extreme Condition Data Extension: When the reliability assessment results indicate that the model's coverage of a certain operating scenario is insufficient, or when an uncovered extreme operating scenario is detected in the current dataset, or when preset rules or event triggering conditions are met, the extreme operating scenario data expansion process is dynamically triggered. Specifically, this includes triggering a vehicle-side active data collection command to supplement scarce operating scenario data, simultaneously synthesizing corresponding extreme operating scenario data through a generative adversarial network, and jointly training it with real data to enhance the model's robustness to extreme conditions. When the assessment result is used as the triggering basis, this assessment can originate from the edge-side assessment of the previous running cycle and / or cloud-side offline batch processing; when the coverage assessment strategy and / or preset events / rules are used as the basis, S2 can be triggered directly without relying on real-time assessment.

[0076] S2.1 Active Sampling Trigger: Calculate the coverage of the currently collected dataset in key environmental dimensions (e.g., temperature-acceleration two-dimensional space); when the number of samples in a certain grid area (e.g., temperature between -20℃ and -10℃ and acceleration > 2.5m / s²) is less than a preset threshold (e.g., 50 samples), send a data collection command to vehicles that meet the conditions via OTA, triggering the vehicle terminal to actively collect data of the uncovered working conditions, and can activate the battery heating system as needed to simulate extreme scenarios such as rapid acceleration at low temperatures.

[0077] S2.2 Adversarial Data Generation: Synthetic data for extreme operating conditions is generated using a Conditional Generative Adversarial Network (WGAN). A generator model is constructed, taking an environmental condition vector as input (e.g., temperature -30℃ to 50℃, slope 0° to 30°, etc.); the output is a simulated battery voltage / current time series curve (sampling rate 10Hz). The discriminator model employs a Wasserstein loss function with a gradient penalty term λ=10 to ensure the stability of the generated data.

[0078] S2.3 Data Mixed Validation: The generated synthetic extreme condition data undergoes quality verification. First, the Dynamic Time Warping (DTW) algorithm is used to calculate the similarity between the synthetic and real curves. Then, the physical plausibility of the synthetic data is verified using a battery mechanism model (e.g., verifying whether the SOC decline rate conforms to the predictions of the Peukert equation). These verifications ensure that the generated extreme condition data is consistent with real data in both statistical characteristics and physical mechanisms. Qualified synthetic data will be mixed with real-world collected data for subsequent model training.

[0079] S3. Dynamic Event Perception and Response: By accessing real-time traffic event data streams, analyzing information such as road congestion levels and accident location coordinates, and combining online learning algorithms to dynamically update the weight parameters of the energy consumption prediction model, the event response latency is controlled within 200ms.

[0080] S3.1 Traffic Data Access: Configure the Apache Flink stream processing engine to connect to real-time traffic API data streams from Gaode or Baidu, and subscribe to traffic event topics (e.g., / traffic / event / {city_id}) via KafkaConnector. Parse and obtain key fields from the JSON formatted event data, including event type (event_type, e.g., 1 for congestion, 2 for accident) and event location (WGS84 coordinates), to achieve real-time road status awareness input.

[0081] S3.2 Event Impact Modeling: A mapping relationship is established between traffic congestion levels and vehicle speed attenuation coefficients to quantify the impact of road conditions on vehicle speed and energy consumption. For example, a speed attenuation coefficient of 0.9 can be predefined for congestion level 1, and 0.7 for congestion level 2. Accident events are defined with corresponding energy consumption impact factors based on their severity. During prediction, the model multiplies the vehicle speed on the corresponding road segment by the aforementioned attenuation coefficients based on the geographical location of the road condition event to correct the input for energy consumption prediction.

[0082] S3.3 Online Weight Update: An online learning algorithm is employed to rapidly update the prediction model in response to the impact of sudden events. For example, the FTRL-Proximal (Follow-The-Regularized-Leader) algorithm is used to gradually adjust the model weights. Parameters such as a learning rate η=0.01 and an L1 regularization coefficient λ1=0.1 are set. An incremental model update is triggered after receiving a certain number (e.g., 10) of new event data, ensuring that the energy consumption prediction model can respond to the latest traffic events within a delay of no more than 200ms. The response information output by the event response prediction unit includes at least: input correction, state / noise parameter adjustment, and / or model update trigger information. The input correction / parameter adjustment is used for immediate adaptation in S4 inference, while the model update trigger information, along with event features, is provided to S6 for incremental training to determine update timing and sample selection. It can also serve as a target condition description for expanding the S2 extreme condition data to constrain the expansion range and sampling strategy.

[0083] S4. Spatiotemporal Dependency Modeling and Adaptive Prediction: A heterogeneous graph of vehicles, road networks, and the environment is constructed based on historical trajectory data and real-time traffic information to capture the spatiotemporal dependencies of vehicle energy consumption, used to predict sudden energy consumption spikes caused by unexpected traffic events. Simultaneously, an adaptive Kalman filter is introduced to correct the model state, dynamically adjusting the noise covariance matrix of the filtering process to adapt to state jumps caused by traffic events. To satisfy data dependencies, the 'fusion feature' used by S4 is the latest available version generated by S5 in the previous prediction cycle; in real-time execution, S5 and S4 run in a pipeline: the fusion feature produced by S5 is used for the next S4 inference. The order of numbering is not limited by the execution sequence.

[0084] S4.1 Heterogeneous Graph Construction: A heterogeneous graph containing different node types and edge relationships is defined to comprehensively represent vehicle, road network, and environmental elements. Specifically, vehicle nodes (attributes include dynamic features such as vehicle speed and acceleration), road network nodes (attributes include static features such as road slope and curvature), and various relationship edges are constructed: such as membership edges between vehicle nodes and road network nodes (indicating that a vehicle is located in a certain road network position), and topological connection edges between road network nodes (indicating the adjacency or connection relationship between roads), etc. Using this heterogeneous graph, vehicle operating states and the road environment can be organically integrated, providing a structured spatiotemporal dependency modeling foundation for energy consumption prediction.

[0085] S4.2 Graph Neural Network Inference: Energy consumption prediction inference is performed using graph neural networks (such as GraphSAGE) on the constructed heterogeneous graph. The GraphSAGE model can set a certain neighbor sampling rate and aggregation strategy. For example, using an 80% sampling rate to select the first-order neighbor nodes of each node, the features of the neighbor nodes are input into LSTM units for encoding, and then the encoding results are averaged (MEAN aggregation) to generate the spatiotemporal feature representation of the central node. Through the message passing and aggregation mechanism of the graph neural network, the model can capture the spatiotemporal correlation between vehicles and the surrounding road network and environment, thereby more accurately predicting sudden changes in vehicle energy consumption caused by events such as sudden congestion.

[0086] S4.3 Variational Bayesian Filtering: An adaptive Kalman filter is introduced into the state estimation of energy consumption prediction to handle model state jumps caused by sudden events. The state transition equation and observation equation of the Kalman filter are designed to characterize the impact of traffic events on vehicle energy consumption, and variational inference is used to dynamically update the probability distribution parameters of the process noise covariance. Specifically, by introducing a prior (e.g., an inverse-Gamma distribution) into the process noise covariance, variational Bayes inference is performed based on the observation data after each traffic event to update the parameters of the inverse-Gamma distribution, achieving adaptive adjustment of process noise and thus improving the stability and accuracy of the energy consumption prediction model in non-stationary environments. The spatiotemporal adaptive prediction module receives and loads the prediction model or its variants issued by the edge-cloud collaborative module as the currently active model; the correction residuals and / or parameter update results generated by filtering are used to feed back and adjust the relation modeling and aggregation strategy of the graph network. S5. Multimodal Feature Extraction and Fusion: Multimodal feature extraction is performed on time-series data of vehicle operation and external environmental data to obtain joint feature vectors that can characterize long-term driving patterns and environmental impacts, thereby improving the adaptability of energy consumption prediction models to different driving behaviors and environmental conditions.

[0087] S5.1 timing embedding coding: For time-series data such as vehicle speed and acceleration, the Informer temporal-transformer model is used to extract long-term temporal pattern features. The input window length of the Informer model is configured to be 60 seconds (60 historical data points are obtained by sampling at 1Hz), and 8 self-attention heads are set to enhance the capture of long-period dependencies. The model encodes the input sequence through a self-attention mechanism and extracts special [CLS] markers as outputs, which serve as temporal feature vectors (e.g., 512-dimensional) representing driving behavior patterns.

[0088] S5.2 Environmental Feature Coding: For external environmental parameters such as temperature, humidity, wind speed, slope, solar radiation intensity, and air pressure, a one-dimensional convolutional residual network (ResNet-1D) is used for feature extraction. The network input layer receives time-synchronized data from the above six environmental dimensions, extracts multi-level features through three residual convolutional modules, and outputs a 256-dimensional environmental feature vector to represent the impact of the current environmental conditions on energy consumption.

[0089] S5.3 Feature Fusion: The obtained temporal feature vector (e.g., 512-dimensional) and environmental feature vector (256-dimensional) are concatenated and fused along their feature dimensions to form a joint feature vector that comprehensively represents both vehicle and environmental states. Subsequently, the concatenated joint features are normalized using a LayerNormalization layer to eliminate scale differences between features of different dimensions, providing stable feature input for subsequent model predictions. Temporal alignment and dimension normalization are performed during the fusion process; furthermore, the environmental encoding unit can be conditionally updated based on key time period indicators provided by the temporal embedding unit to enhance its ability to represent abnormal or extreme conditions. S6. Incremental Model Training and Adaptive Update: Continuously monitor changes in vehicle energy consumption data distribution. When data drift occurs (such as energy consumption shift caused by battery aging), trigger incremental model retraining. Then, use online distillation technology to transfer the knowledge of the full model to the lightweight sub-model, thereby adapting to the evolution of data characteristics during long-term use and the limited computing resources on the vehicle.

[0090] S6.1 Drift Detection: Statistical detection methods are used to monitor changes in the distribution of input data and prediction results. For example, the Page-Hinkley detection algorithm is employed, with an allowable deviation α=0.1 and an alarm threshold λ=30. When the cumulative sum of model prediction errors continuously exceeds the threshold λ, data distribution drift is identified, triggering a model retraining process. This mechanism can promptly identify energy consumption prediction deviations caused by factors such as battery performance degradation and changes in driving habits. The drift detection unit outputs trigger information for model updates, which can be used in parallel or in combination with the trigger information generated by S3.

[0091] S6.2 Online Model Distillation: During model retraining, to balance model accuracy and on-device resource constraints, an online knowledge distillation strategy is employed to train a lightweight model. The fully retrained model (e.g., an LSTM network with 512 hidden layers) is selected as the teacher model, and its predictions are used as a soft objective to guide the learning of the lightweight student model (e.g., a LSTM sub-network pruned to 128 hidden layers). The distillation loss function is defined as the KL divergence between the teacher and student model outputs. By minimizing this loss, the student model parameters are iteratively updated, allowing the student model to maintain prediction performance as close as possible to the teacher model while significantly reducing the number of parameters. Online distillation uses the target prediction model (a candidate version to be loaded and run on the device by S4) as the object, and performs knowledge transfer with reference to at least one control model (a system-maintained reference model for the same task). The resulting validation results and / or parameter updates are used to adjust the triggering strategy and sample selection range of the drift detection unit.

[0092] S7. Edge-Cloud Collaboration and Model Synchronization: Based on the computing power of the vehicle terminal, model versions adapted to different computing loads are dynamically generated and distributed. Differential synchronization technology is used to upload only the differences in model parameters to the cloud to reduce bandwidth consumption and achieve efficient edge-cloud model collaborative updates.

[0093] S7.1 model slice generation: Neural Architecture Search (NAS) is used to automatically generate model slice versions that meet different computing power requirements. First, the NAS search space is defined, including a set of selectable network layer types (e.g., Conv1D convolutional layers, LSTM recurrent layers, Transformer layers, etc.) and a set of network width scaling factors (e.g., model widths of 0.25×, 0.5×, 1×, etc.). Then, a Bayesian optimization algorithm based on TPE (Tree-structured Parzen Estimator) is used to find the optimal model structure within the search space. Model structures of different complexities are selected to correspond to different automotive chip computing power levels. The generated model slice versions are then distributed to the corresponding vehicle terminals to fully utilize the computing power of each terminal.

[0094] S7.2 Differential Model Synchronization: After the model update is completed locally at the edge node, the update results are efficiently uploaded to the cloud via a differential synchronization mechanism. Specifically, the RSync protocol is used for model parameter synchronization, with a block size of 4KB. The MD5 rolling verification algorithm is used to identify differences between the model parameter file and the cloud baseline version, transmitting only the locations of the differing blocks and their patch data. By uploading only the differing parts of the model parameters, the communication overhead between the edge device and the cloud server is significantly reduced, enabling real-time synchronization of model updates. After the cloud side completes the integrity / consistency verification, the spatiotemporal adaptive prediction module loads the target model variant on the edge side and replaces it, taking effect. The parameter differences and / or update information generated during edge-side operation are transmitted back by the differential synchronization unit and aggregated on the cloud side. The aggregation results are used to adjust the model slicing and distribution strategy.

[0095] S8. Result Credibility Assessment and Counterfactual Analysis: The uncertainty of energy consumption prediction results is quantitatively assessed, and scenarios with low confidence are marked. When an extreme operating condition scenario is detected for the first time, a corresponding counterfactual sample is generated to provide traceable evidence and explanation for the prediction results.

[0096] S8.1 Uncertainty Quantification: The Monte Carlo Dropout method is used to assess the uncertainty of the prediction results. During the model inference phase, a certain dropout rate (e.g., 20%) is maintained, and the model output is repeatedly sampled, for example, 100 random dropout inferences are performed. The distribution of the obtained prediction values ​​is statistically analyzed, and a 95% confidence interval is calculated (e.g., the mean of the prediction results μ ± 1.96σ). The confidence interval width is used to determine the reliability of the prediction results. Predictions with large confidence intervals but low confidence are marked, and possible anomalies or unseen conditions are identified, providing a basis for subsequent targeted data expansion.

[0097] S8.2 Counterfactual Sample Generation: A neural process model is introduced to generate counterfactual samples for low-confidence scenarios to aid in result interpretation and model improvement. A neural process encoder-decoder architecture is employed: the encoder is a two-layer fully connected network (MLP, 64 hidden units) that encodes observed environment-energy consumption sample pairs into intermediate representations; the decoder introduces a Gaussian process kernel function (e.g., radial basis function (RBF) kernel, length_scale=1.0) to generate energy consumption predictions for counterfactual samples based on the intermediate representations provided by the encoder. This counterfactual generation model can simulate changes in vehicle energy consumption under slightly different environmental inputs, helping to analyze which factors cause uncertainty in the original predictions and providing clues for further expanding the system to extreme operating condition data.

[0098] The evaluation results are used to determine whether to initiate S2 extreme condition data expansion; the explanatory information (including counterfactual samples and / or comparative scenarios) serves as the target constraint and sample selection basis for S2 data expansion, wherein counterfactual samples can be included in the training dataset as a candidate to supplement training data.

[0099] S9. Model Deployment and Real-Time Prediction Output: The trained lightweight energy consumption prediction model is deployed to the vehicle terminal, and dynamic energy consumption prediction results are output based on the vehicle's real-time input data, providing real-time decision-making reference for the driver or vehicle control system.

[0100] S9.1 Model Quantization: Before deployment, the model is optimized for inference, for example, by using TensorRT to perform INT8 quantization compression. Approximately 1000 typical driving cycles are used as a calibration dataset to quantize and calibrate the model, ensuring that the prediction accuracy of the quantized model meets the requirements (maximum relative error controlled within 1%). The quantized model significantly reduces parameter size and computational load, making it more suitable for real-time inference on in-vehicle terminals.

[0101] S9.2 Real-time Inference: A real-time data processing and inference pipeline is built on the in-vehicle terminal. The collected CAN bus data is averaged using a 20ms sliding window to smooth out instantaneous fluctuations. The vehicle's location obtained from GPS is matched against a map to extract the corresponding road slope information from OpenStreetMap. The preprocessed sensor data is input into the deployed lightweight model, which performs inference cyclically at a fixed 100ms cycle to generate dynamic energy consumption predictions for the vehicle's current state and short-term future. These predictions are then provided to users or uploaded to the cloud via the in-vehicle HMI or communication module.

[0102] Through steps S1 to S9 above, this embodiment utilizes knowledge graph semantic alignment under the federated learning framework to achieve automatic mapping and distributed cleaning of multi-source vehicle data; it combines an active learning strategy to trigger vehicles to actively collect scarce operating condition data, and uses conditional generative adversarial networks to synthesize physically reasonable extreme operating condition data, breaking through the limitation of training data sparsity and significantly improving the model's robustness to extreme conditions; at the same time, it uses a stream processing engine to access traffic event data in real time, constructs a vehicle-road network-environment heterogeneous graph to model complex spatiotemporal dependencies, and integrates variational Bayesian filtering to dynamically and adaptively adjust model noise parameters, combined with online learning algorithms to achieve millisecond-level model weight updates, accurately predicting the instantaneous and continuous impact of sudden events on vehicle energy consumption.

[0103] Example 3 A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to perform the steps of the method described in Example 2.

[0104] Example 4 An electronic device is characterized by comprising a processor and a memory, wherein the memory stores a computer program, and when the program is executed by the processor, the electronic device performs the steps of the method described in Embodiment 2.

[0105] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A new energy vehicle energy consumption prediction system based on big data, characterized in that, comprise each other: a federal data alignment module configured to perform semantic alignment and local cleaning of multi-vehicle and multi-source sensor data under privacy-controlled conditions, and output an aligned feature vector; a multi-modal feature extraction module configured to encode time series signals and environmental signals in the aligned feature vector respectively and output fused features; a dynamic event perception module configured to access a road event data stream and output event features and / or trigger information for model updating; a spatio-temporal adaptive prediction module configured to receive and load a prediction model issued by an edge-cloud collaborative module, and perform vehicle-road network-environment relationship modeling based on the fused features and the event features, and output an energy consumption prediction result; a credibility assessment module configured to quantify the uncertainty of the energy consumption prediction result, and generate an assessment result indicating the prediction reliability; an extreme condition generation module configured to output data expansion instructions to drive supplementary data acquisition and / or conditional data generation based on the assessment result indicating the prediction reliability and / or a data sparse scene determined according to a coverage evaluation strategy, and incorporate the obtained supplementary training data into a training data set to form an expanded training set; an incremental model training module configured to perform retraining or fine-tuning of the prediction model used by the spatio-temporal adaptive prediction module based on the trigger information and the expanded training set, and output an updated model; an edge-cloud collaborative module configured to receive the updated model and issue it to the spatio-temporal adaptive prediction module for replacement and validation.

2. The system of claim 1, wherein, The system comprises vehicle-instance nodes on the edge side and at least one cloud-side node; wherein: The federal data alignment module adopts an edge-cloud collaborative arrangement, the edge side performs noise suppression, abnormality rejection, semantic / format standardization on local multi-source raw data to form an aligned feature vector, and the cloud side aggregates the aligned feature vector and its derived data under privacy-controlled conditions and provides them for use by subsequent modules; The multi-modal feature extraction module, the spatio-temporal adaptive prediction module, and the credibility assessment module are deployed on the edge side; The dynamic event perception module is configured to access a road event data stream on the cloud side and / or the edge side and generate event features and / or trigger information, which is provided to the edge side for use when generated on the cloud side through a communication link; The extreme condition generation module adopts an edge-cloud collaborative arrangement, the cloud side outputs data expansion instructions, and the acquisition execution unit on the edge side implements supplementary data acquisition, and the obtained supplementary training data is incorporated into a training data set to form an expanded training set; The incremental model training module is mainly executed on the cloud side; The edge-cloud collaborative module performs version management and distribution of the updated model on the cloud side, and issues it to the spatio-temporal adaptive prediction module on the edge side for replacement and validation, while recycling edge-side parameter differences and / or update information as needed for cloud-side aggregation.

3. The system of claim 1, wherein, The federal data alignment module comprises a heterogeneous data mapping unit and a federal alignment unit; The heterogeneous data mapping unit preloads and / or dynamically maintains a cross-vehicle field mapping rule library, and the semantic standardization maps the original fields, units, and codes of different vehicles into a standard field set and outputs unified format data by calling the rule library. The federal alignment unit performs time reference alignment, unit normalization, identity association and missing processing based on the uniform format data, and generates an aligned feature vector under a privacy-controlled collaborative mechanism; The field mapping rule library is called by the extreme working condition generation module and the multi-modal feature extraction module, used to maintain the consistency of fields in the subsequent data processing link, and incrementally updated when new vehicle models or fields are added.

4. The system of claim 1, wherein, The extreme working condition generation module includes an active learning sampling unit and an adversarial generation unit; The active learning sampling unit determines a target working condition set and generates acquisition trigger information according to the evaluation result of the prediction reliability and / or according to the coverage evaluation strategy and / or event features, so as to cause the acquisition execution unit to acquire supplementary data under the target working condition; The adversarial generation unit generates synthetic samples corresponding to the target working conditions by taking the working condition descriptions of the target working condition set as generation constraints, and supplements the target working condition samples when the supplementary data is insufficient or unavailable. The outputs of the active learning sampling unit and the adversarial generation unit are merged into supplementary training data after consistency and format specification processing, and are incorporated into the training data set to form an expanded training set.

5. The system of claim 1, wherein, The dynamic event perception module includes a real-time road condition perception unit and an event response prediction unit; The real-time road condition perception unit accesses road event data in a streaming manner and performs parsing, standardization and spatio-temporal alignment to extract event features for energy consumption prediction; The event response prediction unit evaluates the immediate and continuous impact of road events on energy consumption based on the event features, and generates response information for model adaptation, including at least one of: input correction amount, state / noise parameter adjustment amount, and / or model update trigger information; The event features of the real-time road condition perception unit are used as inputs of the event response prediction unit, and the response information output by the event response prediction unit is provided to the spatio-temporal adaptive prediction module for immediate input correction and / or parameter adaptation during inference, and is provided to the incremental model training module for update timing determination and training data selection, and is used as the target working condition description of the extreme working condition generation module to cooperatively complete data expansion.

6. The system of claim 1, wherein, The spatio-temporal adaptive prediction module includes a spatio-temporal graph network unit and an adaptive Kalman filter unit; The spatio-temporal graph network unit is used to build a spatio-temporal relationship representation based on vehicles, road networks and environmental elements, aggregate historical driving information and real-time event features, and output prior estimates and event-sensitive features for energy consumption state estimation; The adaptive Kalman filter unit is used to recursively estimate and correct the energy consumption state based on the prior estimates, the event-sensitive features and the fused features, and adaptively adjust the state transition and noise statistics when an event-driven state change is detected to alleviate the impact of state transition on prediction, and the filtered state estimate is used as the energy consumption prediction result; The correction residual and / or parameter update result generated by the adaptive Kalman filter unit are used to feedback adjust the relationship modeling and aggregation strategy of the spatio-temporal graph network unit.

7. The system of claim 1, wherein, The multi-modal feature extraction module comprises a time series embedding unit and an environment encoding unit. The time series embedding unit is configured to perform representation learning on time series signals related to vehicle operation, and extract long-short term features for describing driving behavior. The environment encoding unit is configured to perform representation learning on environment and road condition parameters to obtain environment representation, and perform time alignment, unit / dimension normalization and feature fusion on the long-short term features and the environment representation, and output fused features. The environment encoding unit can be updated conditionally based on the key time period indication provided by the time series embedding unit to enhance the representation capability of abnormal or extreme working conditions.

8. The system of claim 1, wherein, The incremental model training module comprises a drift detection unit and an online distillation unit. The drift detection unit is configured to identify data distribution and / or concept drift based on running data and prediction bias statistics, and generate trigger information for model updating. The online distillation unit is configured to perform knowledge transfer on the target prediction model on the extended training set after receiving the trigger information, and output an updated model. Wherein: The target prediction model refers to a prediction model candidate version to be loaded and run on the edge side by the spatio-temporal adaptive prediction module. The knowledge transfer is based on the supervision signal provided by at least one reference model, which is a same-task model version maintained by the system. The online distillation unit can determine sample selection and update range, and capacity and / or structure constraints of the target prediction model according to the trigger information, and the verification result and / or parameter update generated thereby are used to feedback and adjust the trigger strategy of the drift detection unit.

9. The system of claim 1, wherein, The edge-cloud collaborative module comprises a model slicing unit and a differential synchronization unit. The model slicing unit is configured to generate and / or select model variants based on the updated model as a baseline, and form variant description information for distribution and replacement according to the calculation, storage and communication constraints and task objectives on the edge side. The differential synchronization unit is configured to perform differential transmission and aggregation between the cloud side and the edge side under privacy-controlled conditions according to the variant description information, synchronize derived update information of the model, and complete verification. Wherein, the model slicing unit determines the target model variant and delivers it to the differential synchronization unit; after verification is completed, the target model variant is loaded by the spatio-temporal adaptive prediction module and replaced to take effect; the update information generated during the running of the edge side is fed back to the differential synchronization unit and aggregated on the cloud side, and the aggregation result is used to adjust the model slicing and distribution strategy.

10. The system of claim 1, wherein, The credibility evaluation module comprises an uncertainty quantification unit and a counterfactual explanation unit. The uncertainty quantification unit is configured to quantify the uncertainty of the energy consumption prediction result, and generate an evaluation result for indicating the prediction reliability. The counterfactual explanation unit is configured to determine the prediction result to be explained based on the evaluation result, construct counterfactual samples and / or contrast scenarios related to the prediction result, and output explanation information. Wherein, the evaluation result is used to determine whether to initiate data expansion, the explanation information is used as a target constraint and sample selection basis for data expansion by the extreme working condition generation module, and the counterfactual samples are incorporated into the training data set as a type of candidate for the supplementary training data. The contrast scenario is used to define a reference condition and / or boundary of difference with the prediction result.

11. A new energy vehicle energy consumption prediction method based on big data, characterized in that, The system implementation according to any one of claims 1-10, comprising: Semantic alignment and local cleaning of multi-vehicle type and multi-source sensor data under privacy-controlled conditions to obtain an aligned feature vector; Encoding and outputting a fusion feature for a time series signal and an environmental signal in the aligned feature vector, respectively; Accessing a road event data stream and generating event features and / or trigger information for model updating; Receiving and loading a prediction model issued by an edge-cloud collaborative module, modeling vehicle-road network-environment relationships based on the fusion feature and the event feature, and outputting an energy consumption prediction result; Quantifying the uncertainty of the energy consumption prediction result to generate an evaluation result for indicating prediction reliability; Based on the evaluation result for indicating prediction reliability and / or a data sparse scenario determined according to an overlay evaluation strategy, outputting data expansion instructions to drive supplementary data acquisition and / or conditional data generation, and incorporating the obtained supplementary training data into a training data set to form an expanded training set; Based on the trigger information and the expanded training set, retraining or fine-tuning the prediction model to obtain an updated model; Through the edge-cloud collaborative module, issuing the updated model to the end side and replacing it to take effect.

12. A computer-readable storage medium, characterized in that, A computer program is stored thereon, which, when executed by a processor, causes the processor to perform the method steps of claim 11.

13. An electronic device, comprising: An electronic device comprising a processor and a memory, wherein the memory stores a computer program, and the program, when executed by the processor, causes the electronic device to perform the method steps of claim 11. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, and the program, when executed by the processor, causes the electronic device to perform the method steps of claim 11.

Images