New energy automobile energy consumption prediction method and system based on artificial intelligence

Abstract

The invention discloses a new energy automobile energy consumption prediction method and system based on artificial intelligence, and belongs to the technical field of new energy automobiles, and the method comprises the steps: S1, multi-source dynamic data collection and fusion, S2, data preprocessing and feature engineering, S3, hybrid model architecture design, S4, dynamic environment adaptation and real-time correction, and S5, prediction result output and optimization. On the basis of predicting the energy consumption of the new energy automobile, the system energy consumption can be assisted to be incorporated into the model, and the method can adapt to a dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle technology, and more specifically, to a method and system for predicting the energy consumption of new energy vehicles based on artificial intelligence. Background Technology

[0002] New energy vehicles refer to automobiles that use unconventional vehicle fuels as their power source or use conventional vehicle fuels but employ new on-board power devices. They integrate advanced technologies in vehicle power control and drive, resulting in automobiles with advanced technical principles, new technologies, and new structures. New energy vehicles include natural gas vehicles, hybrid vehicles, pure electric vehicles, fuel cell electric vehicles, solar-powered vehicles, and other new energy vehicles.

[0003] Traditional dynamic models require precise acquisition of parameters such as vehicle mass and drag coefficient, which is difficult to collect in batches in practice, leading to prediction bias. At the same time, fixed driving cycles cannot simulate real complex road conditions, ignore the impact of dynamic scenarios such as rapid acceleration and sudden slope changes, and do not integrate external data such as temperature, humidity, and wind speed. This results in large errors in air conditioning power consumption estimation, insufficient low-temperature correction of battery internal resistance, and distorted winter range prediction. Summary of the Invention

[0004] In view of the problems existing in the prior art, the purpose of this invention is to provide a method and system for predicting the energy consumption of new energy vehicles based on artificial intelligence. In addition to predicting the energy consumption of new energy vehicles, this invention can also assist the system in incorporating energy consumption into the model and can adapt to dynamic environments.

[0005] To solve the above problems, the present invention adopts the following technical solution: An artificial intelligence-based method for predicting the energy consumption of new energy vehicles, comprising: S1. Multi-source dynamic data acquisition and fusion; S2, Data Preprocessing and Feature Engineering; S3, Hybrid Model Architecture Design; S4. Dynamic environment adaptation and real-time correction; S5. Prediction result output and optimization.

[0006] As a preferred embodiment of the present invention, step S1, multi-source dynamic data acquisition and fusion, includes: Real-time acquisition of core driving data: The vehicle speed, acceleration, motor power and battery status are collected in real time through the CAN vehicle sensor bus. At the same time, the path gradient, curve curvature and real-time traffic congestion prediction information are obtained by combining high-precision maps. Auxiliary system energy consumption monitoring: Monitor the real-time power consumption of the auxiliary system, which includes air conditioning, multimedia, lighting and seat heating. The power consumption of the air conditioning includes cooling / heating power, ambient temperature and humidity and operating time. Charging behavior characteristic monitoring: integrates battery remaining capacity at the start / end of charging, charging temperature curve, peak current and single charge quantity parameters; External environmental data access: Access meteorological API to obtain environmental temperature, humidity and wind speed parameters, and at the same time correct the air conditioning energy consumption model; Driving behavior and historical data recording: Records driving habits, historical energy consumption records, and vehicle static attributes.

[0007] As a preferred embodiment of the present invention, step S2, data preprocessing and feature engineering, includes: Missing value imputation: For missing battery parameters, a random forest regression model is used, combined with Pearson correlation coefficient to select highly correlated features for joint imputation; Logical rule correction: Abnormal vehicle speed is logically reset to zero, and incorrect charging status marking is directly rejected; Outlier removal: Remove data with vehicle speed greater than 0 and invalid vehicle status flags that conflict with logic when the vehicle is parked and charging. For continuous variables such as current and voltage, remove data that deviates too much from the mean based on the 3σ principle. Trip segmentation: Trip segments are divided according to vehicle status, charging status, and time continuity, and segments with a sample size of less than 20 are deleted; Remove invalid data: Data segments in which the current / vehicle speed did not return to zero after the vehicle was turned off were directly deleted; Work segment subdivision: The travel segment is divided into work segments of 1 kilometer each, reflecting the periodic changes in speed or acceleration within the road segment; Dynamic driving characteristics: Quantify driving stability, mark rapid acceleration or deceleration behavior as high energy consumption characteristics, calculate the sine component of the slope angle based on high-precision map elevation data, generate the slope resistance coefficient in combination with vehicle weight, introduce the recovery efficiency coefficient, and correct the energy consumption value in frequent start-stop scenarios. Auxiliary system features: Predicts peak power consumption of air conditioners under different temperatures and humidity levels based on long short-term memory networks, and estimates energy consumption based on historical switching probabilities combined with real-time running time; Battery degradation compensation features: Calculate the battery capacity degradation rate based on the cumulative number of charge and discharge cycles and usage time, and dynamically adjust the battery internal resistance coefficient under low temperature conditions; Feature selection and dimensionality reduction: Pearson correlation coefficient is used to screen highly relevant features and remove redundant features. Feature importance assessment: Feature importance is calculated based on a random forest model, while retaining vehicle speed, gradient, and SOH battery health status; Data standardization and partitioning: The continuous features of vehicle speed and temperature are scaled to the [0,1] interval, and the features of acceleration and current that conform to the normal distribution are normalized by mean and variance.

[0008] Dataset partitioning: The dataset is divided into a 70% training set, a 15% validation set, and a 15% test set based on timestamps to prevent future data leaks.

[0009] As a preferred embodiment of the present invention, the formula used in step S2 includes: The formula for imputing missing values ​​in random forest is: ,in For the random forest regression model, These are the characteristic matrices excluding the j-th column; The outlier correction formula is: ,in The standard deviation of the feature is expressed in units identical to the original feature. The mean of the features is expressed in units identical to those of the original features. The formula for the slope resistance coefficient is: Where m is the vehicle weight in kg, and g is the gravitational acceleration 9.8 m / s². 2 , Slope angle, output force The formula for converting to energy consumption coefficient is: , t represents vehicle speed in km / h and time in hours. The formula for the dynamic model of air conditioner energy consumption is: Where k is the thermal conductivity coefficient and its unit is kW / ℃. The unit is the air conditioner's operating time in hours. The ambient temperature is expressed in °C. The formula for correcting the internal resistance of a battery at low temperatures is as follows: ,in The value is the internal resistance at 25℃ in Ω, and k is the attenuation coefficient with a value of 0.03 / ℃. The actual battery temperature is expressed in °C.

[0010] As a preferred embodiment of the present invention, the normalization formula for the input features is: ; The Z-Score standardization formula is: ; The logarithmic normalization formula is ,in It is a tiny constant; The feature importance screening formula is: .

[0011] As a preferred embodiment of the present invention, the hybrid model architecture design in step S3 includes: Model component selection and function allocation: embed vehicle dynamics equations, quantify basic driving energy consumption, process static features and low-dimensional time-series features through the extreme gradient boosting algorithm, and capture long-term dependencies of vehicle speed and acceleration time-series data through a long short-term memory network; Auxiliary system energy consumption monitoring; predicting operating time and peak power consumption under different temperatures and humidity conditions using long short-term memory networks; Environment Adaptation Module: Processes global dependencies of high-dimensional environmental features through a Transformer encoder, and extracts local pattern recognition from road condition images or weather radar data through local features; Feature-level fusion: The temporal features output by the Long Short-Term Memory network are concatenated with the environmental features encoded by the Transformer and input into the fully connected layer. The weights of the outputs of each module are dynamically allocated through a self-attention mechanism. Predictive-level fusion: The residuals of the Long Short-Term Memory (LSTM) network predicting the physical model are superimposed and then denormalized. The prediction results of the Extreme Gradient Boosting (ELS) algorithm and the Lightweight Gradient Boosting (LTG) algorithm are used as input features for the LSM network, and then the network is learned a second time through the meta-model.

[0012] As a preferred embodiment of the present invention, the formula used in step S3, hybrid model architecture design, is as follows: The joint output formula is ,in , For deep network output, For the Sigmoid function; The UMoE attention routing formula is: ,in For routing probability, Output for expert networks.

[0013] As a preferred embodiment of the present invention, step S4, dynamic environment adaptation and real-time correction, includes: Sliding window alignment: Edge computing nodes are used to perform time interpolation on asynchronous data to ensure timing consistency; Semantic mapping: converting road slope into energy consumption coefficient; Probabilistic boundary modeling: Establishing the relationship between sensor detection probability and visibility. ,when When the value is less than 0.8, an early warning is triggered. The environmental parameters and system status are combined to output a dimensionless risk score R∈[0,1]. Tiered response mechanism: When R < 0.3, it is a Level 1 warning and only records data; when 0.3 ≤ R < 0.7, it is a Level 2 warning and the sensor scanning frequency is increased; when R ≥ 0.7, it is a Level 3 warning and the system is forced to switch to safe mode. Image comparison: Detecting road occlusions using the YOLO model triggers incremental training; Data filtering: retain features that show significant changes and remove noisy data; Online model update: Fine-tune the weights of the long short-term memory network layers, and decay the learning rate to 10% of the original value to avoid overfitting. If the false alarm rate of the new model increases, it will automatically roll back to the stable version. Dynamic parameter correction: The battery internal resistance is dynamically adjusted according to temperature T, and the power consumption of the air conditioner is predicted in real time based on temperature and humidity. Path deviation correction: Solving a quadratic programming problem with speed constraints. With steering angle The formula is ,in Q is the weight matrix, and the speed profile is updated every 100ms to adapt to sudden road conditions; Dynamic resource allocation: High-time-sensitive tasks are handled by the vehicle-mounted lightweight model, while complex predictions are handed over to the cloud. Expert models are dynamically allocated through the MoE gating network, with sparsity constraints ensuring that the number of activated experts is ≤3.

[0014] As a preferred embodiment of the present invention, step S5, prediction result output and optimization, includes: Prediction results output: Convert the energy consumption values ​​predicted by the model into structured data, with the unit being kWh / km, and label it with timestamps and vehicle ID metadata; Persistent storage: Results are saved to a database for easy subsequent analysis and traceability; Visualization: Compare predicted and actual values ​​using line charts, mark key operating points, and display real-time energy consumption trends and deviation heatmaps in FineBI to assist decision-making.

[0015] API service encapsulation: Deploy the model as a microservice to receive real-time vehicle data and return prediction results; Prediction result evaluation: Calculate the root mean square error and the mean absolute percentage error; the target root mean square error is ≤1.5kWh. Residual analysis: Plot the residual distribution to identify systematic biases; Reliability testing: Verify model stability in stratified environments of >35℃ high temperature and <-10℃ low temperature; Sensitivity testing: Identify key features of disturbances and observe the predicted fluctuation amplitude; Practicality assessment: Analyze the impact of the prediction results on the error in the estimated driving range; Real-time feedback: The in-vehicle terminal displays the remaining range with dynamic adjustment values, increasing user trust; Hyperparameter tuning: Grid search was used to adjust the number of layers and learning rate parameters of the long short-term memory network, and hyperparameters were sampled independently for different sequence elements within a batch; Model structure optimization: Introduce dynamic features or lagged variables; Hybrid architecture improvement: The extreme gradient boosting algorithm is used to predict the residuals of the dynamic equations, and the weight ratio of the Transformer module is increased when the environment changes abruptly.

[0016] A new energy vehicle energy consumption prediction method and system based on artificial intelligence includes: a multimodal data acquisition module, a data synchronization and alignment module, a feature construction module, a path segmentation module, a historical association module, a hybrid prediction model module, a result output module, a path planning module, and a model optimization module; The multimodal data acquisition module includes vehicle-mounted sensors, LiDAR, V2X (Vehicle-to-Everything) connectivity, and a meteorological API. The vehicle-mounted sensors collect real-time data on vehicle speed, acceleration, battery temperature, voltage, current, and tire pressure. The LiDAR acquires slope angle and road geometry. The V2X connectivity receives dynamic traffic information. The meteorological API integrates temperature, humidity, and wind speed data. The data synchronization and alignment module interpolates asynchronous data through edge nodes, unifies timestamps, and converts slope angle into an energy consumption coefficient. The data preprocessing module fills in missing values, unifies dimensions, and normalizes the data. The feature construction module calculates the slope drag coefficient, regenerative braking efficiency, and historical average vehicle speed. The path segmentation module is used to divide road segments according to speed limits or to perform secondary segmentation according to gradient changes, along with acceleration variance. The historical association module is used to match historical energy consumption of the same road segments, and to match similar road segments through cosine similarity when no match is found. The hybrid prediction model module is used to calculate basic driving energy consumption, learn long-term dependencies of vehicle speed sequences through a long short-term memory network, and process static features through the extreme gradient boosting algorithm. The result output module is used for structured output and visualization processing. The path planning module is used for dynamic path optimization and updates the remaining range in real time when deviating from the navigation, triggering energy-saving suggestions or replanning charging points. The model optimization module is used to fine-tune the weights of the long short-term memory network every 5 minutes to adapt to sudden environmental changes.

[0017] Compared with the prior art, the advantages of this invention are: (1) This invention enhances comprehensiveness through multi-source fusion, covering multi-dimensional information such as driving data, environmental data, and driving behavior, eliminating single data bias, quantifying key factors such as slope resistance and air conditioning power consumption through feature engineering, providing high-value input for the model, balancing accuracy and generalization in the hybrid model, ensuring the scientific nature of basic energy consumption calculations in the physical model, capturing temporal and environmental features in the long short-term memory network, and compensating for errors in complex scenarios through the residual correction mechanism.

[0018] (2) This invention improves the prediction robustness during environmental changes through dynamic weight allocation, reduces RMSE to 1.27 kWh, achieves a 50ms response time with a lightweight model at the edge, recalculates complex scenarios in the cloud, updates parameters through incremental learning, dynamically switches temperature stratification models to adapt to all climate scenarios, improves driving safety through a risk grading mechanism, assists users in decision-making through visualization results, drives model iteration through residual analysis, and updates the model through A / B testing, forming a closed loop and reducing users' range anxiety. Attached Figure Description

[0019] Figure 1 This is a flowchart of an artificial intelligence-based method for predicting the energy consumption of new energy vehicles according to the present invention. Figure 2 This is a block diagram of an artificial intelligence-based energy consumption prediction system for new energy vehicles according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] Example:

[0022] Please see Figure 1 A new energy vehicle energy consumption prediction method based on artificial intelligence includes: S1, multi-source dynamic data acquisition and fusion; S2, Data Preprocessing and Feature Engineering; S3, Hybrid Model Architecture Design; S4. Dynamic environment adaptation and real-time correction; S5. Prediction result output and optimization.

[0023] In specific embodiments of the present invention, comprehensiveness is enhanced through multi-source fusion, covering multi-dimensional information such as driving data, environmental data, and driving behavior, eliminating single data biases. Feature engineering quantifies key factors such as slope resistance and air conditioning power consumption, providing high-value inputs for the model. The hybrid model balances accuracy and generalization, the physical model ensures the scientific nature of basic energy consumption calculations, long short-term memory networks capture temporal and environmental features respectively, residual correction mechanisms compensate for errors in complex scenarios, dynamic weight allocation improves prediction robustness during environmental changes, RMSE is reduced to 1.27 kWh, the lightweight model at the edge achieves a 50ms response, complex scenarios are recalculated in the cloud, parameters are updated incrementally, temperature-layered models are dynamically switched to adapt to all climate scenarios, risk grading mechanisms improve driving safety, visualized results assist user decision-making, residual analysis drives model iteration, A / B testing is rolled up to form a closed loop, and user range anxiety is reduced.

[0024] Specifically, step S1, multi-source dynamic data acquisition and fusion, includes: Real-time acquisition of core driving data: The vehicle speed, acceleration, motor power and battery status are collected in real time through the CAN vehicle sensor bus. At the same time, the path gradient, curve curvature and real-time traffic congestion prediction information are obtained by combining high-precision maps. Auxiliary system energy consumption monitoring: Monitor the real-time power consumption of auxiliary systems, which include air conditioning, multimedia, lighting and seat heating. Air conditioning power consumption includes cooling / heating power, ambient temperature and humidity and operating time. Charging behavior characteristic monitoring: integrates battery remaining capacity at the start / end of charging, charging temperature curve, peak current and single charge quantity parameters; External environmental data access: Access meteorological API to obtain environmental temperature, humidity and wind speed parameters, and at the same time correct the air conditioning energy consumption model; Driving behavior and historical data recording: Records driving habits, historical energy consumption records, and vehicle static attributes.

[0025] In a specific embodiment of the present invention, core driving data such as vehicle speed, acceleration, and motor power are collected in real time via the vehicle-mounted CAN bus. Combined with high-precision maps, path gradient and traffic congestion information are obtained. Simultaneously, the power consumption of the air conditioning and multimedia assistance systems is monitored. Charging behavior characteristics and environmental data are integrated, and driving habits and vehicle static attributes are recorded to achieve multi-dimensional data fusion, eliminate data silos, provide comprehensive input for energy consumption prediction, cover multiple factors such as driving, environment, and vehicle status, improve the integrity and reliability of the prediction basis data, and avoid prediction distortion caused by the bias of a single data source.

[0026] Specifically, step S2, data preprocessing and feature engineering, includes: Missing value imputation: For missing battery parameters, a random forest regression model is used, combined with Pearson correlation coefficient to select highly correlated features for joint imputation; Logical rule correction: Abnormal vehicle speed is logically reset to zero, and incorrect charging status marking is directly rejected; Outlier removal: Remove data with vehicle speed greater than 0 and invalid vehicle status flags that conflict with logic when the vehicle is parked and charging. For continuous variables such as current and voltage, remove data that deviates too much from the mean based on the 3σ principle. Trip segmentation: Trip segments are divided according to vehicle status, charging status, and time continuity, and segments with a sample size of less than 20 are deleted; Remove invalid data: Data segments in which the current / vehicle speed did not return to zero after the vehicle was turned off were directly deleted; Work segment subdivision: The travel segment is divided into work segments of 1 kilometer each, reflecting the periodic changes in speed or acceleration within the road segment; Dynamic driving characteristics: Quantify driving stability, mark rapid acceleration or deceleration behavior as high energy consumption characteristics, calculate the sine component of the slope angle based on high-precision map elevation data, generate the slope resistance coefficient in combination with vehicle weight, introduce the recovery efficiency coefficient, and correct the energy consumption value in frequent start-stop scenarios. Auxiliary system features: Predicts peak power consumption of air conditioners under different temperatures and humidity levels based on long short-term memory networks, and estimates energy consumption based on historical switching probabilities combined with real-time running time; Battery degradation compensation features: Calculate the battery capacity degradation rate based on the cumulative number of charge and discharge cycles and usage time, and dynamically adjust the battery internal resistance coefficient under low temperature conditions; Feature selection and dimensionality reduction: Pearson correlation coefficient is used to screen highly relevant features and remove redundant features. Feature importance assessment: Feature importance is calculated based on a random forest model, while retaining vehicle speed, gradient, and SOH battery health status; Data standardization and partitioning: The continuous features of vehicle speed and temperature are scaled to the [0,1] interval, and the features of acceleration and current that conform to the normal distribution are normalized by mean and variance.

[0027] Dataset partitioning: The dataset is divided into a 70% training set, a 15% validation set, and a 15% test set based on timestamps to prevent future data leaks. The formulas used in step S2 include: The formula for imputing missing values ​​in random forest is: ,in For the random forest regression model, These are the characteristic matrices excluding the j-th column; The outlier correction formula is: ,in The standard deviation of the feature is expressed in units identical to the original feature. The mean of the features is expressed in units identical to those of the original features. The formula for the slope resistance coefficient is: Where m is the vehicle weight in kg, and g is the gravitational acceleration 9.8 m / s². 2 , Slope angle, output force The formula for converting to energy consumption coefficient is: , t represents vehicle speed in km / h and time in hours. The formula for the dynamic model of air conditioner energy consumption is: Where k is the thermal conductivity coefficient and its unit is kW / ℃. The unit is the air conditioner's operating time in hours. The ambient temperature is expressed in °C. The formula for correcting the internal resistance of a battery at low temperatures is as follows: ,in The value is the internal resistance at 25℃ in Ω, and k is the attenuation coefficient with a value of 0.03 / ℃. The actual battery temperature is expressed in °C. The normalization formula for the input features is: ; The Z-Score standardization formula is: ; The logarithmic normalization formula is ,in It is a tiny constant; The feature importance screening formula is:

[0028] In a specific embodiment of the present invention, random forest regression is used to fill in missing battery parameter values, logical rules are used to correct abnormal vehicle speeds, the 3σ principle is used to remove abnormal current / voltage data, the journey is segmented according to vehicle status, the working condition segments are divided, the gradient resistance coefficient and air conditioning energy consumption features are dynamically extracted, and key features are screened by Pearson correlation coefficient and random forest to improve data quality and feature effectiveness, quantify driving behavior and environmental impact, reduce noise interference, provide high-value input for the model, and avoid overfitting caused by redundant features.

[0029] Specifically, step S3, the hybrid model architecture design, includes: Model component selection and function allocation: embed vehicle dynamics equations, quantify basic driving energy consumption, process static features and low-dimensional time-series features through the extreme gradient boosting algorithm, and capture long-term dependencies of vehicle speed and acceleration time-series data through a long short-term memory network; Auxiliary system energy consumption monitoring; predicting operating time and peak power consumption under different temperatures and humidity conditions using long short-term memory networks; Environment Adaptation Module: Processes global dependencies of high-dimensional environmental features through a Transformer encoder, and extracts local pattern recognition from road condition images or weather radar data through local features; Feature-level fusion: The temporal features output by the Long Short-Term Memory network are concatenated with the environmental features encoded by the Transformer and input into the fully connected layer. The weights of the outputs of each module are dynamically allocated through a self-attention mechanism. Predictive-level fusion: The residuals of the Long Short-Term Memory Network predicting the physical model are superimposed and then denormalized. The prediction results of the Extreme Gradient Boosting Algorithm and the Lightweight Gradient Boosting Machine are used as the input features of the Long Short-Term Memory Network, and then learned twice through the meta-model. The formula used in step S3, the hybrid model architecture design: The joint output formula is ,in , For deep network output, For the Sigmoid function; The UMoE attention routing formula is: ,in For routing probability, Output for expert networks.

[0030] In a specific embodiment of the present invention, the physical model calculates the basic driving energy consumption, the extreme gradient boosting algorithm processes static features, and the long short-term memory network captures long-term dependencies of the vehicle speed sequence; the Transformer encodes environmental features and concatenates them with the output of the long short-term memory network through a self-attention mechanism; the residual correction module uses the extreme gradient boosting algorithm to predict the physical model residuals, and outputs the final energy consumption value after superposition. This integrates the advantages of physical laws and data-driven approaches, with the long short-term memory network solving temporal dependencies, the Transformer enhancing global modeling of environmental features, and the residual correction improving the prediction accuracy of complex scenarios, reducing RMSE to 1.27 kWh.

[0031] Specifically, step S4, dynamic environment adaptation and real-time correction, includes: Sliding window alignment: Edge computing nodes are used to perform time interpolation on asynchronous data to ensure timing consistency; Semantic mapping: converting road slope into energy consumption coefficient; Probabilistic boundary modeling: Establishing the relationship between sensor detection probability and visibility. ,when When the value is less than 0.8, an early warning is triggered. The environmental parameters and system status are combined to output a dimensionless risk score R∈[0,1]. Tiered response mechanism: When R < 0.3, it is a Level 1 warning and only records data; when 0.3 ≤ R < 0.7, it is a Level 2 warning and the sensor scanning frequency is increased; when R ≥ 0.7, it is a Level 3 warning and the system is forced to switch to safe mode. Image comparison: Detecting road occlusions using the YOLO model triggers incremental training; Data filtering: retain features that show significant changes and remove noisy data; Online model update: Fine-tune the weights of the long short-term memory network layers, and decay the learning rate to 10% of the original value to avoid overfitting. If the false alarm rate of the new model increases, it will automatically roll back to the stable version. Dynamic parameter correction: The battery internal resistance is dynamically adjusted according to temperature T, and the power consumption of the air conditioner is predicted in real time based on temperature and humidity. Path deviation correction: Solving a quadratic programming problem with speed constraints. With steering angle The formula is ,in Q is the weight matrix, and the speed profile is updated every 100ms to adapt to sudden road conditions; Dynamic resource allocation: High-time-sensitive tasks are handled by the vehicle-mounted lightweight model, while complex predictions are handed over to the cloud. Expert models are dynamically allocated through the MoE gating network, with sparsity constraints ensuring that the number of activated experts is ≤3.

[0032] In a specific embodiment of the present invention, edge nodes align asynchronous data timestamps, convert slope angles into energy consumption coefficients in real time, establish a visibility detection probability model, trigger a graded response with risk score R∈[0,1], trigger incremental training with YOLO detection of road obstructions, fine-tune the weights of the long short-term memory network, dynamically correct battery internal resistance with temperature, achieve real-time response to sudden environmental changes, adapt to road condition changes with incremental learning, improve accuracy in low-temperature scenarios with dynamic internal resistance correction, and ensure driving safety with a risk grading mechanism.

[0033] Specifically, step S5, the prediction result output and optimization, includes: Prediction results output: Convert the energy consumption values ​​predicted by the model into structured data, with the unit being kWh / km, and label it with timestamps and vehicle ID metadata; Persistent storage: Results are saved to a database for easy subsequent analysis and traceability; Visualization: Compare predicted and actual values ​​using line charts, mark key operating points, and display real-time energy consumption trends and deviation heatmaps in FineBI to assist decision-making.

[0034] API service encapsulation: Deploy the model as a microservice to receive real-time vehicle data and return prediction results; Prediction result evaluation: Calculate the root mean square error and the mean absolute percentage error; the target root mean square error is ≤1.5kWh. Residual analysis: Plot the residual distribution to identify systematic biases; Reliability testing: Verify model stability in stratified environments of >35℃ high temperature and <-10℃ low temperature; Sensitivity testing: Identify key features of disturbances and observe the predicted fluctuation amplitude; Practicality assessment: Analyze the impact of the prediction results on the error in the estimated driving range; Real-time feedback: The in-vehicle terminal displays the remaining range with dynamic adjustment values, increasing user trust; Hyperparameter tuning: Grid search was used to adjust the number of layers and learning rate parameters of the long short-term memory network, and hyperparameters were sampled independently for different sequence elements within a batch; Model structure optimization: Introduce dynamic features or lagged variables; Hybrid architecture improvement: The extreme gradient boosting algorithm is used to predict the residuals of the dynamic equations, and the weight ratio of the Transformer module is increased when the environment changes abruptly.

[0035] In a specific embodiment of the present invention, the output energy consumption value is stored as structured data; the comparison between predicted and actual values ​​and the deviation heatmap are visualized; RMSE and MAPE are calculated to evaluate accuracy; residual analysis identifies system deviations; high and low temperature stratification verifies the robustness of the model; grid search optimizes the hyperparameters of the long short-term memory network; dynamic features are introduced to improve generalization ability; multidimensional evaluation of model performance is provided; visualization assists user decision-making; hyperparameter tuning and feature enhancement continuously improve prediction accuracy; and extreme environment verification ensures practicality.

[0036] Please see Figure 2 A new energy vehicle energy consumption prediction system based on artificial intelligence includes: a multimodal data acquisition module, a data synchronization and alignment module, a feature construction module, a path segmentation module, a historical association module, a hybrid prediction model module, a result output module, a path planning module, and a model optimization module; The multimodal data acquisition module includes vehicle-mounted sensors, LiDAR, V2X (Vehicle-to-Everything) connectivity, and a meteorological API. Vehicle-mounted sensors collect real-time data on vehicle speed, acceleration, battery temperature, voltage, current, and tire pressure. LiDAR acquires slope angle and road geometry. V2X connects to the vehicle network to receive dynamic traffic information. The meteorological API provides access to temperature, humidity, and wind speed data. The data synchronization and alignment module interpolates asynchronous data through edge nodes, unifies timestamps, and converts slope angle into an energy consumption coefficient. The data preprocessing module fills in missing values, unifies dimensions, and normalizes the data. The feature construction module calculates the slope drag coefficient, regenerative braking efficiency, historical average vehicle speed, and acceleration. The speed variance and path segmentation modules are used to divide road segments according to speed limits or to perform secondary segmentation based on gradient changes. The historical association module is used to match historical energy consumption of the same road segment, and to match similar road segments using cosine similarity when no match is found. The hybrid prediction model module is used to calculate basic driving energy consumption, learn long-term dependencies of vehicle speed sequences through a long short-term memory network, and process static features through the extreme gradient boosting algorithm. The result output module is used for structured output and visualization. The path planning module is used for dynamic path optimization and updates the remaining range in real time when deviating from the navigation, triggering energy-saving suggestions or replanning charging points. The model optimization module is used to fine-tune the weights of the long short-term memory network every 5 minutes to adapt to sudden environmental changes.

[0037] In a specific embodiment of the present invention, end-to-end automation reduces the need for manual intervention, real-time battery life updates alleviate users' range anxiety, and improve accuracy.

[0038] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the energy consumption of new energy vehicles based on artificial intelligence, characterized in that, include: S1. Multi-source dynamic data acquisition and fusion; S2, Data Preprocessing and Feature Engineering; S3, Hybrid Model Architecture Design; S4. Dynamic environment adaptation and real-time correction; S5. Prediction result output and optimization.

2. The method for predicting the energy consumption of new energy vehicles based on artificial intelligence according to claim 1, characterized in that, Step S1, multi-source dynamic data acquisition and fusion, includes: Real-time acquisition of core driving data: The vehicle speed, acceleration, motor power and battery status are collected in real time through the CAN vehicle sensor bus. At the same time, the path gradient, curve curvature and real-time traffic congestion prediction information are obtained by combining high-precision map. Auxiliary system energy consumption monitoring: Monitor the real-time power consumption of the auxiliary system, which includes air conditioning, multimedia, lighting and seat heating. The power consumption of the air conditioning includes cooling / heating power, ambient temperature and humidity and operating time. Charging behavior characteristic monitoring: integrates battery remaining capacity at the start / end of charging, charging temperature curve, peak current and single charge quantity parameters; External environmental data access: Access meteorological API to obtain environmental temperature, humidity and wind speed parameters, and at the same time correct the air conditioning energy consumption model; Driving behavior and historical data recording: Records driving habits, historical energy consumption records, and vehicle static attributes.

3. The method for predicting the energy consumption of new energy vehicles based on artificial intelligence according to claim 2, characterized in that, Step S2, data preprocessing and feature engineering, includes: Missing value imputation: For missing battery parameters, a random forest regression model is used, combined with Pearson correlation coefficient to select highly correlated features for joint imputation; Logical rule correction: Abnormal vehicle speed is logically reset to zero, and incorrect charging status marking is directly rejected; Outlier removal: Remove data with vehicle speed greater than 0 and invalid vehicle status flags that conflict with logic when the vehicle is parked and charging. For continuous variables such as current and voltage, remove data that deviates too much from the mean based on the 3σ principle. Trip segmentation: Trip segments are divided according to vehicle status, charging status, and time continuity, and segments with a sample size of less than 20 are deleted; Remove invalid data: Data segments in which the current / vehicle speed did not return to zero after the vehicle was turned off were directly deleted; Work segment subdivision: The travel segment is divided into work segments of 1 kilometer each, reflecting the periodic changes in speed or acceleration within the road segment; Dynamic driving characteristics: Quantify driving stability, mark rapid acceleration or deceleration behavior as high energy consumption characteristics, calculate the sine component of the slope angle based on high-precision map elevation data, generate the slope resistance coefficient in combination with vehicle weight, introduce the recovery efficiency coefficient, and correct the energy consumption value in frequent start-stop scenarios. Auxiliary system features: Predicts peak power consumption of air conditioners under different temperatures and humidity levels based on long short-term memory networks, and estimates energy consumption based on historical switching probabilities combined with real-time running time; Battery degradation compensation features: Calculate the battery capacity degradation rate based on the cumulative number of charge and discharge cycles and usage time, and dynamically adjust the battery internal resistance coefficient under low temperature conditions; Feature selection and dimensionality reduction: Pearson correlation coefficient is used to screen highly relevant features and remove redundant features. Feature importance assessment: Feature importance is calculated based on a random forest model, while retaining vehicle speed, gradient, and SOH battery health status; Data standardization and partitioning: Continuous features of vehicle speed and temperature are scaled to the [0,1] interval, and the mean and variance of features of acceleration and current that conform to a normal distribution are normalized; Dataset partitioning: The dataset is divided into a 70% training set, a 15% validation set, and a 15% test set based on timestamps to prevent future data leaks.

4. The method for predicting the energy consumption of new energy vehicles based on artificial intelligence according to claim 3, characterized in that, The formulas used in step S2 include: The formula for imputing missing values ​​in random forest is: ,in For the random forest regression model, These are the characteristic matrices excluding the j-th column; The outlier correction formula is: ,in The standard deviation of the feature is expressed in units identical to the original feature. The mean of the features is expressed in units identical to those of the original features. The formula for the slope resistance coefficient is: Where m is the vehicle weight in kg, and g is the gravitational acceleration 9.8 m / s². 2 , Slope angle, output force The formula for converting to energy consumption coefficient is: , t represents vehicle speed in km / h and time in hours. The formula for the dynamic model of air conditioner energy consumption is: Where k is the thermal conductivity coefficient and its unit is kW / ℃. The unit is the air conditioner's operating time in hours. The ambient temperature is expressed in °C. The formula for correcting the internal resistance of a battery at low temperatures is as follows: ,in The value is the internal resistance at 25℃ in Ω, and k is the attenuation coefficient with a value of 0.03 / ℃. The actual battery temperature is expressed in °C.

5. The method for predicting the energy consumption of new energy vehicles based on artificial intelligence according to claim 4, characterized in that, The normalization formula for the input features is: ; The Z-Score standardization formula is: The logarithmic normalization formula is ,in It is a tiny constant; The feature importance screening formula is: .

6. The method for predicting the energy consumption of new energy vehicles based on artificial intelligence according to claim 5, characterized in that, The hybrid model architecture design in step S3 includes: Model component selection and function allocation: embed vehicle dynamics equations, quantify basic driving energy consumption, process static features and low-dimensional time-series features through the extreme gradient boosting algorithm, and capture long-term dependencies of vehicle speed and acceleration time-series data through a long short-term memory network; Auxiliary system energy consumption monitoring; predicting operating time and peak power consumption under different temperatures and humidity conditions using long short-term memory networks; Environment Adaptation Module: Processes global dependencies of high-dimensional environmental features through a Transformer encoder, and extracts local pattern recognition from road condition images or weather radar data through local features; Feature-level fusion: The temporal features output by the Long Short-Term Memory network are concatenated with the environmental features encoded by the Transformer and input into the fully connected layer. The weights of the outputs of each module are dynamically allocated through a self-attention mechanism. Predictive-level fusion: The residuals of the Long Short-Term Memory (LSTM) network predicting the physical model are superimposed and then denormalized. The prediction results of the Extreme Gradient Boosting (ELS) algorithm and the Lightweight Gradient Boosting (LTG) algorithm are used as input features for the LSM network, and then the network is learned a second time through the meta-model.

7. The method for predicting the energy consumption of new energy vehicles based on artificial intelligence according to claim 6, characterized in that, The formula used in step S3, the hybrid model architecture design, is as follows: The joint output formula is, where, For deep network output, For the Sigmoid function; The UMoE attention routing formula is: ,in For routing probability, Output for expert networks.

8. The method for predicting the energy consumption of new energy vehicles based on artificial intelligence according to claim 7, characterized in that, Step S4, dynamic environment adaptation and real-time correction, includes: Sliding window alignment: Edge computing nodes are used to perform time interpolation on asynchronous data to ensure timing consistency; Semantic mapping: converting road slope into energy consumption coefficient; Probabilistic boundary modeling: Establishing the relationship between sensor detection probability and visibility. ,when When the value is less than 0.8, an early warning is triggered. The environmental parameters and system status are combined to output a dimensionless risk score R∈[0,1]. Tiered response mechanism: When R < 0.3, it is a Level 1 warning and only records data; when 0.3 ≤ R < 0.7, it is a Level 2 warning and the sensor scanning frequency is increased; when R ≥ 0.7, it is a Level 3 warning and the system is forced to switch to safe mode. Image comparison: Detecting road occlusions using the YOLO model triggers incremental training; Data filtering: retain features that show significant changes and remove noisy data; Online model update: Fine-tune the weights of the long short-term memory network layers, and decay the learning rate to 10% of the original value to avoid overfitting. If the false alarm rate of the new model increases, it will automatically roll back to the stable version. Dynamic parameter correction: The battery internal resistance is dynamically adjusted according to temperature T, and the power consumption of the air conditioner is predicted in real time based on temperature and humidity. Path deviation correction: Solving a quadratic programming problem with speed constraints. With steering angle The formula is ,in Q is the weight matrix, and the speed profile is updated every 100ms to adapt to sudden road conditions; Dynamic resource allocation: High-time-sensitive tasks are handled by the vehicle-mounted lightweight model, while complex predictions are handed over to the cloud. Expert models are dynamically allocated through the MoE gating network, with sparsity constraints ensuring that the number of activated experts is ≤3.

9. The method for predicting the energy consumption of new energy vehicles based on artificial intelligence according to claim 8, characterized in that, The prediction result output and optimization in step S5 includes: Prediction results output: Convert the energy consumption values ​​predicted by the model into structured data, with the unit being kWh / km, and label it with timestamps and vehicle ID metadata; Persistent storage: Results are saved to a database for easy subsequent analysis and traceability; Visualization: Compare predicted and actual values ​​using line charts, mark key operating points, and display real-time energy consumption trends and deviation heatmaps in the FineBI tool to assist decision-making; API service encapsulation: Deploy the model as a microservice to receive real-time vehicle data and return prediction results; Prediction result evaluation: Calculate the root mean square error and the mean absolute percentage error; the target root mean square error is ≤1.5kWh. Residual analysis: Plot the residual distribution to identify systematic biases; Reliability testing: Verify model stability in stratified environments of >35℃ high temperature and <-10℃ low temperature; Sensitivity testing: Identify key features of disturbances and observe the predicted fluctuation amplitude; Practicality assessment: Analyze the impact of the prediction results on the error in the estimated driving range; Real-time feedback: The in-vehicle terminal displays the remaining range with dynamic adjustment values, increasing user trust; Hyperparameter tuning: Grid search was used to adjust the number of layers and learning rate parameters of the long short-term memory network, and hyperparameters were sampled independently for different sequence elements within a batch; Model structure optimization: Introduce dynamic features or lagged variables; Hybrid architecture improvement: The extreme gradient boosting algorithm is used to predict the residuals of the dynamic equations, and the weight ratio of the Transformer module is increased when the environment changes abruptly.

10. An artificial intelligence-based energy consumption prediction system for new energy vehicles, applied to the artificial intelligence-based energy consumption prediction method for new energy vehicles as described in claim 1, characterized in that, include: The system includes a multimodal data acquisition module, a data synchronization and alignment module, a feature construction module, a path segmentation module, a historical correlation module, a hybrid prediction model module, a result output module, a path planning module, and a model optimization module. The multimodal data acquisition module includes vehicle-mounted sensors, LiDAR, V2X (Vehicle-to-Everything) connectivity, and a meteorological API. The vehicle-mounted sensors collect real-time data on vehicle speed, acceleration, battery temperature, voltage, current, and tire pressure. The LiDAR acquires slope angle and road geometry. The V2X connectivity receives dynamic traffic information. The meteorological API integrates temperature, humidity, and wind speed data. The data synchronization and alignment module interpolates asynchronous data through edge nodes, unifies timestamps, and converts slope angle into an energy consumption coefficient. The data preprocessing module fills in missing values, unifies dimensions, and normalizes the data. The feature construction module calculates the slope drag coefficient, regenerative braking efficiency, and historical average vehicle speed. The path segmentation module is used to divide road segments according to speed limits or to perform secondary segmentation according to gradient changes, along with acceleration variance. The historical association module is used to match historical energy consumption of the same road segments, and to match similar road segments through cosine similarity when no match is found. The hybrid prediction model module is used to calculate basic driving energy consumption, learn long-term dependencies of vehicle speed sequences through a long short-term memory network, and process static features through the extreme gradient boosting algorithm. The result output module is used for structured output and visualization processing. The path planning module is used for dynamic path optimization and updates the remaining range in real time when deviating from the navigation, triggering energy-saving suggestions or replanning charging points. The model optimization module is used to fine-tune the weights of the long short-term memory network every 5 minutes to adapt to sudden environmental changes.

Images