A new energy vehicle power prediction monitoring method and system based on the Internet of Things

Abstract

The application discloses a new energy vehicle power prediction monitoring method and system based on the Internet of Things, and belongs to the field of new energy vehicle monitoring and control, and comprises the following steps: through collecting new energy vehicle multi-source data and processing, a training sample set for power prediction is formed and input into a residual prediction model, basic power is calculated according to dynamics, residual prediction results for correcting the basic power are output, and total predicted power is obtained; actual data is acquired in the initial stage of a new vehicle going online, a calibration layer parameter is adaptively updated based on the error of actual power and total predicted power, calibrated predicted power results are output, and data drift degree is monitored. Through training the model by using a small sample data set, the application achieves the effects of reducing the number of training samples, reducing the model training cost and operation and maintenance cost, and solves the problems that in the prior art, when automobile power is predicted based on big data, massive data is strongly dependent, engineering cost is high, and it is difficult to quickly adapt to new vehicles or new regions.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle monitoring and control technology, and in particular to a new energy vehicle power prediction and monitoring method and system based on the Internet of Things. Background Technology

[0002] With the large-scale promotion and application of new energy vehicles in the market, the overall energy management, range prediction, charging planning, and fleet operation scheduling of most existing electric vehicles all rely on power prediction and monitoring during vehicle operation. Existing new energy vehicle power prediction and monitoring systems generally equip vehicles with onboard terminals, which upload vehicle operating parameters and regional environmental parameters to a cloud platform via the Internet of Things (IoT). Combined with big data and machine learning models, these systems predict vehicle power, providing crucial data support for vehicle operation decisions.

[0003] In real-time prediction and optimized control of vehicle power, on the one hand, while ensuring driving safety and power response requirements, energy consumption per unit mileage is reduced, improving vehicle range and energy efficiency; on the other hand, it effectively suppresses instantaneous high-power surges and frequent power fluctuations, reducing electrothermal stress and mechanical fatigue in motors and other power electronic drive devices, thereby effectively extending the service life of key components. Furthermore, by utilizing single-vehicle power predictions uploaded via the Internet of Things, centralized monitoring and statistical analysis of the operating status of multiple vehicles can be performed, providing reliable data support for fleet operation scheduling, charging plan customization, and the diagnosis and maintenance of vehicles with abnormal energy consumption. This achieves comprehensive optimization of vehicle operation economy, reliability, and maintainability at the system level.

[0004] For example, Chinese invention patent CN109558988B discloses a method and system for predicting the energy consumption of electric vehicles based on big data fusion, which includes: firstly, decoding, cleaning, and interpolating the historical vehicle operation data; then dividing the processed vehicle operation data into driving status data and charging status data; and finally fusing the driving status data with the corresponding environmental information data to obtain the set of energy consumption level prediction parameters required for model training and energy consumption prediction.

[0005] The above-mentioned technology has at least the following technical problems:

[0006] Most existing solutions for vehicle power prediction require massive amounts of historical data from the vehicle or region to train the prediction model in order to improve prediction accuracy. This makes it difficult to quickly adapt to power prediction based on new vehicles or regions, resulting in excessively high engineering costs. Furthermore, while existing technologies employ small-sample adaptive strategies to train models adapted to new vehicles or regions, the lack of quantitative evaluation methods for the information content and representativeness of small samples makes it difficult to ensure that the minimum number of samples covers as many typical operating conditions as possible. This leads to overfitting when training models with small-sample adaptive datasets.

[0007] In addition, existing technologies continuously monitor vehicles during actual operation and identify data with significant deviations as drift data, which are then packaged and uploaded to the cloud for retraining. However, due to the lack of a mechanism for decomposing and attributing the source of drift after drift detection, the retraining scope is either too large or too small, resulting in the inability to accurately repair data and wasting computing resources. Summary of the Invention

[0008] To address the technical problems of existing technologies that heavily rely on massive amounts of historical data for vehicle power prediction and cannot quickly adapt to new vehicles or new regions, this invention provides an Internet of Things-based method and system for predicting and monitoring the power of new energy vehicles. The technical solution is as follows:

[0009] On the one hand, an IoT-based method for predicting and monitoring the power of new energy vehicles is provided, including:

[0010] Multi-source data of new energy vehicles is collected through vehicle terminals, edge nodes, or cloud platforms. This multi-source data includes at least one or more of the following: vehicle operating status data, vehicle parameter data, road and environmental condition data, and vehicle or area identification data. The multi-source data is cleaned, aligned, missing data is removed, and normalized and fused to form a training sample set for power prediction. Based on the training sample set, a base power value is calculated according to the vehicle's energy consumption mechanism or dynamic mechanism to characterize the base power component of the vehicle under the current operating conditions. The training sample set is input into a residual prediction model, which outputs residual prediction results to correct the base power value. The base power value and the residual prediction results are fused to generate the total predicted power of the new energy vehicle. During the trial operation phase of the vehicle's online operation, actual power data for the corresponding time period is acquired. Before the accumulated amount of actual power data reaches a preset update condition, the calibration layer parameters are adaptively updated based on the error between the actual power data and the total predicted power to output a calibrated predicted power result, while keeping the core parameters of the residual prediction model unchanged. The calibrated predicted power result is output, along with monitoring information generated from the error, for power prediction drift judgment or operation strategy adjustment.

[0011] On the other hand, an Internet of Things-based power prediction and monitoring system for new energy vehicles is provided, including:

[0012] The multi-source data acquisition and preprocessing module is used to collect multi-source data from new energy vehicles through vehicle terminals, edge nodes or cloud platforms, and to perform data cleaning, outlier removal, missing value completion, time alignment and normalization fusion processing on the multi-source data to form a training sample set with complete structure, consistent time sequence and controllable quality.

[0013] The basic power calculation and feature construction module is used to calculate the basic power value of the vehicle under the current operating conditions based on the feature parameters in the training sample set and according to the vehicle energy consumption mechanism model or dynamic model. Based on this, a prediction condition vector related to the residual is constructed for the input of the subsequent residual prediction model.

[0014] The residual prediction and total power generation module is used to input the prediction condition vector into the residual prediction model to obtain the residual prediction result used to correct the base power value. The base power value and the residual prediction result are fused and superimposed to generate the total predicted power of the new energy vehicle under the current operating conditions. At the same time, the module outputs the predicted power sequence with model version and time stamp.

[0015] The hierarchical parameter online update module is used to acquire the actual power data for the corresponding time period during the trial operation phase of the vehicle. Before the cumulative amount of actual power data reaches the preset update condition, the calibration layer parameters are adaptively updated based on the error between the actual power data and the total predicted power to output the calibrated predicted power result, while keeping the backbone parameters of the residual prediction model unchanged.

[0016] The predicted power output and drift monitoring module is used to output calibrated predicted power results and monitoring information generated by errors, which is used for power prediction drift judgment or operation strategy adjustment.

[0017] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0018] 1. The IoT-based power prediction and monitoring method for new energy vehicles provided by this invention effectively filters out low-quality samples that lack key signals, have abnormal noise, or have extreme residuals by introducing sample quality scoring and information content scoring mechanisms. It prioritizes the allocation of limited small sample budgets to samples that have higher update value for the pre-trained model and can expand the coverage of regional operating conditions. This reduces the interference of abnormal data on the small sample adaptation process from the data source, thereby reducing the number of training samples for new vehicles or new regions. As a result, it can achieve rapid adaptation of new vehicles or new regions without relying on massive amounts of new vehicle or new region data, significantly reducing the cost of model training and data collection. It effectively solves the problem that the training of power prediction models in the prior art requires massive amounts of historical data and cannot be quickly promoted and applied.

[0019] 2. This invention introduces a hierarchical calibration module and constructs drift metrics for different factors, freezing most model parameters and assigning only a small number of adaptation parameters to a single vehicle. This significantly reduces the scale of small-sample adaptation parameters. When a new vehicle or region is deployed, only a very small number of parameters need to be updated to achieve rapid and personalized model correction. At the same time, it realizes automatic identification and attribution of drift sources, effectively reducing the risk of overfitting under limited sample conditions. This provides a scientific basis for selecting local model fine-tuning according to needs, avoiding blind global training. It not only reduces the computing resources and engineering investment required for model maintenance, but also lowers the adaptation threshold and operation and maintenance costs when new vehicles or regions are deployed. This is conducive to rapid promotion and long-term stable operation in large-scale fleet and multi-region deployment scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of an IoT-based power prediction and monitoring method for new energy vehicles provided in this application embodiment;

[0022] Figure 2 A flowchart illustrating the method for constructing a small sample dataset as provided in this application embodiment;

[0023] Figure 3 Flowchart of the online data distribution drift monitoring and triggering algorithm provided in the embodiments of this application;

[0024] Figure 4 A schematic diagram of the structure of the Internet of Things-based new energy vehicle power prediction and monitoring system provided in the embodiments of this application;

[0025] Figure 5 This is a diagram of the core model architecture for residual prediction. Detailed Implementation

[0026] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that embodiments of the present disclosure may be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure.

[0027] It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. In the description of the embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects.

[0028] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0029] This invention provides a power prediction and monitoring method for new energy vehicles based on the Internet of Things, such as... Figure 1 The diagram shown is a flowchart of an IoT-based power prediction and monitoring method for new energy vehicles provided in this application embodiment. The method includes the following steps:

[0030] S1: Collect historical sample data for training the residual prediction backbone model. The collected data includes historical vehicle operating status data, historical vehicle parameter data, road and environmental condition data, and vehicle or area identification data. The specific data collection scope is as follows:

[0031] Historical vehicle operating status data, including vehicle speed and actual power P during vehicle operation. 真实 The data includes: motor current and voltage, air conditioning load, ambient temperature, slope, and road condition codes during actual operation; historical vehicle parameter data, including vehicle model, rated motor power, battery capacity, equipment weight, and tire specifications; road and environmental condition data, as well as vehicle or area identification data, including typical road network slope distribution, climate zones, and average vehicle speed patterns within the area.

[0032] S2: After performing data cleaning, anomaly handling, and time-series alignment on the collected data, a training sample set with complete structure, consistent time series, and controllable quality is formed. The specific processing steps are as follows:

[0033] S21: Data Cleaning. Import the collected raw sample data, perform unit conversion and format standardization on each field to ensure that the units are the same for easy subsequent calculations, remove non-numeric characters (such as spaces, Chinese units, etc.), retain the record with more complete fields for duplicate records, and directly remove records with missing key fields.

[0034] S22: Anomaly Handling. Based on the above steps, each sample data is screened according to its physical reasonable range and physical relationship. If an anomaly is found to be a short-term isolated point, it is replaced by interpolation with normal points before and after it; if an anomaly is a long-term continuous anomaly, the sample is directly removed.

[0035] S23: Time alignment. Establish a unified timeline and align data from different frequencies to this timeline. On the unified timeline, align fields from different sources according to time to form a complete record at any given moment. If data with missing key fields still exists at a certain moment, it should be removed directly.

[0036] S3: Based on vehicle dynamics, motor efficiency curves, and auxiliary power consumption experience models, and using physical formulas or empirical models, the "baseline power value" is calculated based on collected historical vehicle parameters. This baseline power value represents the power that the vehicle physically uses at a fixed point in time, including vehicle drive power, air conditioning power, and accessory power. It should be noted that actual calculations include more than just these three power components. This embodiment only illustrates these three power sub-items; technicians can choose to calculate appropriate sub-items to better suit real-world scenarios. All physical units in this embodiment use the International System of Units (SI). The specific calculation process is as follows.

[0037] A1: Vehicle driving power. First, calculate the resistance encountered by the wheels at various points, using the following formula:

[0038]

[0039]

[0040]

[0041]

[0042] Where m is the total vehicle mass (kg), including passengers and cargo, and g is the gravitational acceleration (the value of g varies depending on the geographical location of the vehicle's operating area during specific deployments, and does not affect the overall process and implementation effect). v 'a' is the vehicle's instantaneous speed (m / s), and 'a' is the vehicle's longitudinal acceleration (m / s²). 2 ), f r ρ is the rolling resistance coefficient (provided by the tire manufacturer), ρ is the air density (calculated from temperature and air pressure provided by the meteorological bureau), and C is the rolling resistance coefficient.d It is the drag coefficient (provided by the vehicle manufacturer), A f The frontal area of ​​a vehicle (m²) 2 (Provided by the vehicle manufacturer), i is the road slope (in this embodiment, a sine approximation is used; when i > 0, it is an uphill section, and when i < 0, it is a downhill section), and δ is the equivalent rotational mass coefficient (provided by the vehicle manufacturer).

[0043] Based on the resistance calculated above, the total net force required from the wheels can be obtained, as shown in the following formula:

[0044]

[0045] At this time, the mechanical drive power at the wheel is:

[0046]

[0047] From this, we can obtain the motor power, that is, the vehicle drive power:

[0048]

[0049] Where, η 驱动 It refers to the efficiency of the electric drive system, which is generally obtained by looking up an efficiency map based on torque and speed.

[0050] A2: Air Conditioner Power. Air conditioner power cannot be accurately expressed using a purely physical closed-form formula; therefore, an empirical model is used. This embodiment provides a simple and implementable empirical form for regressing air conditioner power as a reference. In actual operation, technicians can adjust it according to the actual situation. The formula is as follows:

[0051]

[0052]

[0053] in, v The instantaneous vehicle speed (m / s) is α0, α1, α2, and α3, which are parameters obtained through calibration or fitting. T 环境 It is the current ambient temperature, T 车内设定 This is the current set temperature of the car's air conditioning. T is the difference between the ambient temperature and the set temperature of the vehicle's air conditioning.

[0054] A3: Accessory Power. Accessory power can be approximated as a "constant + superposition of several discrete states". An embodiment of this invention provides a typical formula for calculating accessory power as an example, as follows:

[0055]

[0056] Among them, P 车灯 P音响 P 基础附件 A specific typical value can be obtained from the manufacturer's manual.

[0057] Add the power calculated from A1, A2, and A3 above to obtain the decomposed skeleton power, as shown in the following formula:

[0058]

[0059] P 预测 The power decomposition skeleton prediction power is calculated on the t-th sample. Based on this, the residual r between the actual power and the skeleton prediction power is calculated, as shown in the following formula:

[0060]

[0061] Simultaneously, features related to the residuals are encoded into a prediction condition vector x. These features include vehicle parameter encoding, current vehicle operating status, environmental variables, regional features, and historical short-time window statistics. For each collected data point, residual calculation and prediction condition vector encoding are performed to obtain several training pairs (x, r), which are used as training data to drive the residual prediction model.

[0062] After providing the various power calculation formulas in A1 to A3, this embodiment further specifies the hierarchical management and update process for the parameters involved in the formulas.

[0063] Relevant parameters are categorized into three levels: vehicle-level, model-level, and platform-level, and added to their respective parameter libraries. For example, the vehicle-level parameter library stores variable parameters specific to a single vehicle, such as instantaneous speed, current estimated gross weight, battery capacity and health status, and the vehicle's location. The model-level parameter library stores parameters related to a specific model but generally fixed within a fleet of vehicles of the same model, such as the vehicle's nominal curb weight, frontal area and drag coefficient, tire specifications, and rolling resistance coefficient range. The platform-level parameter library stores general physical constants and default empirical values. When a new vehicle or region is added, the system executes a parameter initialization process and retrieves valid values ​​for each parameter according to its priority. During operation, for updates to real-time observable parameters (such as vehicle speed, acceleration, ambient temperature, and air pressure), the system does not persist the version; it only retains short-term history for power calculation and residual statistics. For physical parameters that are difficult to measure directly (such as rolling resistance coefficient and air conditioning empirical parameters), the system performs online correction; the specific correction process is detailed in step S3.

[0064] S4: Select a lightweight deep network model, using the prediction condition vector x as input and the residual r as output, and train it using machine learning methods to obtain the residual prediction backbone model, denoted as f. θ (x), where θ represents the total number of parameters in the model that affect the residual judgment. Its architecture diagram is as follows: Figure 5 As shown. It should be noted that, in addition to the model exemplified in the embodiments of the present invention, other time series prediction models based on machine learning can also be selected for training. In practice, technicians can choose different model structures for training according to their needs.

[0065] S5: The sum of the predicted power and the residual prediction backbone is the total predicted power, which can be expressed by the following formula:

[0066]

[0067] Where P is the total predicted power. In subsequent steps, since P 预测 (x) is calculated from physical formulas, that is, the part of power that actually exists in a physical sense, therefore P 预测 (x) is partially frozen, only affecting f. θ (x) undergoes changes. It's important to note that freezing doesn't mean no change; rather, it means fixing the calculation formulas and calibration parameters of the physical model branches without updating them. At each sampling time, P is calculated based on the current input feature x. 预测 (x), and only for the residual model f θ The parameters of (x) are trained or fine-tuned to make the total predicted power P = P 预测 (x)+f θ (x) Adapt to new vehicles or new areas.

[0068] S6: When a new vehicle or new area is added, collect parameters for the new vehicle, including vehicle model, rated power of the motor, battery capacity, equipment weight, and tire specifications; collect characteristics of the new area, including typical road network slope distribution, climate zones, and average vehicle speed patterns within the area. The collected parameters and characteristics, along with real-time operational data, are then input into the total predicted power calculation formula in S5. Specifically, at sampling time point t, the skeleton power is first calculated, and then the prediction condition vector x is constructed. t Input the residual model and output r t The two are added together to obtain the preliminary total predicted power P. At this point, the preliminary predicted power can be obtained without the need for historical operating data from new vehicles or new areas.

[0069] Since the predicted power in S5 is derived from the residual prediction backbone model trained on other vehicle data, it may not be applicable to newly deployed vehicles or regions. Therefore, a small dataset (i.e., a small sample dataset) is constructed by collecting a small number of high-quality samples, such as... Figure 2 As shown, for the residual prediction backbone model f θ (x) Perform further training to improve the accuracy of predictions.

[0070] After a new vehicle or new region is launched, the predicted power in S5 is used for a trial period. Simultaneously, the system collects all operational data from this trial period as a candidate adaptation sample pool, denoted as D={(x i ,y i )}, where x i Let y be the prediction condition vector for the i-th sample. i This refers to the actual power during operation. It should be noted that the operational data mentioned in this embodiment for a trial period refers to all operational data within one week of vehicle operation. This is only an example and reference; technicians can adjust the trial period according to actual conditions during actual operation, and this embodiment does not impose any restrictions on this.

[0071] For each sample data x in the above-mentioned adapted sample pool, a quality score is calculated. The quality score must include at least signal completeness, signal stability, and outlier residuals. The specific evaluation process is as follows:

[0072] B1: Signal completeness, i.e., the proportion of missing key signals related to vehicle power. Key signals related to vehicle power include parameters such as instantaneous speed, current, voltage, air conditioning load, and ambient temperature, which can be substituted into the formula in S1 for calculation. Let k be the ranking of the key signals in the sample, i.e., the k-th key signal; K be the number of key signals the sample should possess; and t be a specific time point within the time window T. The selection of the time window T is random, and technicians can adjust and select it according to needs and actual conditions during actual operation. The signal completeness is scored and denoted as Q. comp :

[0073]

[0074] B2: Signal stability, i.e., the degree of change of the signal within the sampling time window. Similarly, for the k-th key signal, the difference is calculated within the window, with the specific formula as follows:

[0075]

[0076] Among them, s k (t) represents the value of the k-th key signal at time t, s k (t-1) represents the value of the k-th key signal at the sampling point at time t-1. s k (t) represents the degree of sampling change within adjacent sampling time points. A differential energy E is defined. k :

[0077]

[0078] E kThe larger the value, the less stable the signal is within that sampling time window.

[0079] The signal stability is scored and denoted as Q. stab :

[0080]

[0081] Where, ω k Represents the weights, satisfying ∑ω k =1 condition is enough.

[0082] B3: Abnormal residuals, meaning that under the current operating conditions, the residual value of this sample is significantly larger than that of most samples or exhibits an anomalous direction. For the i-th sample, a total predicted power P can be obtained from the calculation formula in S1. i Meanwhile, the actual collected power P true Subtracting them yields the residual r. i :

[0083]

[0084] Establish a stable sample set. Select sample data from the most recent T minutes as a candidate stable sample set. The sample data includes instantaneous vehicle speed, SOC, gradient, predicted power, actual power, and residuals. Retain samples from the candidate stable sample set that do not have large speed fluctuations or direction changes to form the stable sample set.

[0085] The standardized residual is calculated for each sample within the sampling time window, using the following formula:

[0086]

[0087] Among them, z i Let i represent the standardized residual of the i-th sample. To represent a very small real number, to prevent the denominator from being 0, μ r σ represents the average residual over a stable sample set. r This represents the standard deviation of the residuals on this stable sample set. i The larger the value, the more the sample residual deviates from the normal level. An anomaly detection threshold can be manually set, for example, when |z| < 0.05. i When | > 3, the sample point is judged as a strong outlier and needs to be removed; when 2 < |z i When |≤3, the sample point is judged as a moderate outlier and is reduced in weight during selection. Outlier residuals are scored and denoted as Q. resid :

[0088]

[0089] Where, λ kRepresents the weights, satisfying ∑λ k =1 condition is enough.

[0090] Based on the above steps, the sample quality score Q(x) is obtained:

[0091]

[0092] Here, α, β, and γ represent the weights of the signal completeness score, signal stability score, and abnormal residual score, respectively, ensuring that α+β+γ=1.

[0093] It should be noted that the above scoring calculation formula is for reference only. Technicians can adjust the calculation process according to different actual application scenarios. The embodiments of the present invention only illustrate the above calculation process and do not impose specific constraints on the formula.

[0094] A regional work condition feature vector z(x) is constructed, including average vehicle speed range, acceleration fluctuation, braking frequency, ambient temperature range, weather type, road type, slope range, and curve density. K-means clustering is used to divide the work conditions into several "typical work condition clusters." The number of clusters is adaptively determined based on the number of candidate samples; that is, different numbers of clusters are selected depending on the number of samples, ensuring that each cluster contains no less than a preset lower limit of samples, avoiding overly fine cluster division that results in too few samples in some work condition clusters. After clustering, each candidate sample in this space is assigned a work condition cluster ID. This step ensures that the small sample fitting dataset D can cover all high-frequency work condition clusters as much as possible, rather than being concentrated on a single work condition, avoiding the problem of model underfitting due to concentrated sample features.

[0095] f, which has already been calculated in S4 θ Under the premise of (x), the gradient information is calculated for each candidate sample in the candidate adaptation sample pool, and the specific formula is as follows:

[0096]

[0097] Wherein, L(f) θ (x),y) are about f θ The Lagrangian functions of G(x) and y. The larger G(x) (i.e., gradient information content), the higher the value of the sample for updating the current model. A larger sample may be added to a small sample dataset.

[0098] Similarly, the nearest neighbor distance d(x) between each candidate sample and the currently selected sample set S on the working condition vector is calculated using the following formula:

[0099]

[0100] Where S is the selected sample set, denoted as S={x1,x2,x3,...}, and x' represents sample x in the sample set S. The larger d(x) is, the more different the working conditions represented by the sample is from those represented by the existing samples in the current sample candidate set, and the more it can cover the new working condition area.

[0101] Based on the sample quality score Q(x), gradient information G(x), and nearest neighbor distance d(x) calculated above, the sample information I(x) is obtained by weighted summation, as shown in the following formula:

[0102]

[0103] Among them, η1, η2, and η3 are weighting coefficients, which are set by technical personnel according to the actual situation, and the ratio is generally 1:1:1.

[0104] Based on the information I(x) calculated above, a greedy sampling algorithm is driven to select the most suitable 100-300 data points from the candidate sample pool D as a small sample adaptation dataset for training the residual prediction backbone model f. θ (x). The specific implementation process of the greedy sampling algorithm is as follows:

[0105] Given a sample budget N, first initialize the small sample fitting dataset, making D adapt The set is empty. Next, from the working condition clusters, prioritize selecting clusters that are not yet covered, and within these clusters, select the sample with the largest I(x) and add it to D. adapt Repeat this step until all "high-frequency operating condition clusters" are covered by at least one sample or the number of samples reaches N / 2. If the number of samples is still less than N, continue to add D to the entire candidate set in descending order of I(x). adapt Continue until N is reached. Finally, output D. adapt This serves as a small sample adaptation dataset. The specific value of N can be any natural number between 30 and 500. This embodiment does not impose constraints on the selection of the sample size; technicians can choose an appropriate value based on the actual situation.

[0106] Because f obtained in S4 θ (x) can only output the result, and relevant technical personnel cannot determine the specific factors causing the residual in actual applications. Based on this, in f θ By inserting a calibration layer into the output layer of (x), the model update effect can be achieved by adjusting only a few parameters. The specific implementation process is as follows:

[0107] Model f θ The first L-1 layers of (x) transform the input x to obtain an intermediate feature vector, denoted as h:

[0108]

[0109] In the final layer, h is transformed to obtain the output f. θ (x):

[0110]

[0111] Where W is the weight matrix, i.e., the weights of all connections from each feature to each output in the output layer, arranged in rows and columns; b is the bias layer; and θ1 is the weights and bias parameters of each hidden layer in the backbone network for residual prediction. These three are collectively referred to as the backbone parameters, which remain unchanged during model updates. This is a standard output formula for a neural network. After the output layer, a small function g(h;Φ) is added as a calibration layer, i.e.:

[0112]

[0113] Since the intermediate feature vector h has already been decomposed, θ1, W, and b are all frozen during model training. Only the calibration layer parameters Φ need to be adjusted to update the entire model, significantly reducing engineering costs and computational overhead. The calibration layer parameters Φ do not include θ1, W, b, or hyperparameters such as the learning rate and sample quality score weights, which are not learned through backpropagation.

[0114] To clarify f θ The specific layering in (x) facilitates subsequent monitoring processes. In this embodiment, g(h;Φ) is further refined, and the parameter Φ is divided into three layers: platform layer, vehicle model layer, and vehicle layer. The specific formula is as follows:

[0115]

[0116] Here, Φ1 represents all parameters in the vehicle-level calibration layer, Φ2 represents all parameters in the model-level calibration layer, and Φ3 represents all parameters in the platform-level calibration layer. Φ1, Φ2, and Φ3 at least include the weight matrix and bias terms of the corresponding calibration submodules. During routine operation and maintenance, vehicle-level parameter Φ1 is updated first, and model-level parameter Φ2 is updated when necessary, while Φ3 remains largely unchanged. The power prediction backbone model after adding the calibration layer is retrained using a small-sample adaptation dataset, and the training result is used as the updated model and applied to the vehicle.

[0117] It should be noted that the hierarchical classification in this embodiment is for reference only. In actual use, technicians can make different hierarchical classifications according to the actual situation. This embodiment does not limit the hierarchical classification standard.

[0118] S7: After the above model is applied to new vehicles and new areas, the vehicles or areas are put into practical use. During vehicle operation, the system collects vehicle operating status and area characteristics in real time, constructs a prediction error sequence based on the collected data, and monitors information such as the difference between the prediction error sequence and key input features and historical training distributions in real time. Figure 3 As shown.

[0119] The prediction error sequence is a time series that represents the error between the actual power of a vehicle operating within a certain time window and the predicted power obtained from the prediction power model. It is denoted as:

[0120]

[0121] To determine the error situation within the time window T, the average deviation is calculated using the following formula:

[0122]

[0123] If the average bias > 0 for an extended period during vehicle operation, it indicates that the model frequently underestimates the actual power, meaning the actual power is greater than the predicted power; conversely, it indicates that the model frequently overestimates the actual power.

[0124] The difference between key input features and historical training distributions can be quantified using a formula. First, calculate the difference index for each individual feature, denoted as D. dep The specific formula is as follows:

[0125]

[0126] Where, μ 实际输入 μ represents the mean of the actual feature inputs within this time window. 历史训练 σ represents the mean of the historical training features. 历史训练 The standard deviation representing the characteristics of historical training Represent a very small real number to prevent the denominator from being 0. D dep The larger the value, the more severe the average level drift of this feature.

[0127] The above calculation is performed on each key input feature to obtain several D values. dep For these several Ds dep We perform a weighted summation to obtain a drift metric score:

[0128]

[0129] Where, ω j These are weights set manually and adjusted according to actual conditions. Generally, features with greater influence, such as vehicle speed and temperature, are given greater weights. This embodiment does not impose any constraints on these weights.

[0130] A drift threshold is manually set. If D 漂移 If the threshold is exceeded for n consecutive time windows, the model is considered to have significant drift on that vehicle. All data within the time windows exhibiting significant drift are packaged and uploaded to the cloud-based residual prediction backbone model for retraining. After model training, the model is deployed to the vehicle for implementation. To ensure version consistency during model updates, the vehicle appends the current model version number to each prediction result and monitoring data point during residual prediction. The cloud uses this field to verify the consistency between the actual running version on the vehicle and the issued configuration. Simultaneously, the model version label corresponding to each sample is retained when constructing the retraining dataset to avoid label pollution caused by mixing different version data.

[0131] Example 2: Due to the lack of a mechanism for decomposing and attributing the source of drift after drift detection, the retraining range may be too large or too small, resulting in either excessive waste of resources or an inability to accurately repair the key factors causing drift. Therefore, to solve this problem, Example 2 of this invention, based on Example 1, maintains the statistical characteristics of three channels in addition to the original overall error monitoring, and adjusts different parameters according to the different needs of the three channels, further reducing computational overhead. The specific implementation steps are as follows:

[0132] Based on the formula in Example 1 S6, in f θ A calibration layer is introduced on the output side of (x), as shown in the following formula:

[0133]

[0134] During online operation, in addition to the overall error monitoring in Example 1, statistical characteristics are maintained for three channels: vehicle condition, vehicle type, and platform. Vehicle condition channel characteristics include battery usage depth distribution, SOC range, and charge / discharge rate; vehicle type channel characteristics include vehicle curb weight, tire specifications, electric drive system efficiency map, and default accessory power of the vehicle type; platform channel characteristics include external temperature distribution, seasonal labels, weather type, and extreme weather ratio.

[0135] Within a sliding time window T, referring to the calculation formula for the drift metric S7 in Example 1, the drift metric for each channel is calculated and denoted as D. 车辆级漂移 D 车型级漂移 D 平台级漂移 These correspond to vehicle condition channel characteristics, vehicle model channel characteristics, and platform channel characteristics, respectively. Similarly, a threshold is artificially set for each of the three drift metrics, denoted as T. 车辆级 T 车型级 T 平台级 Based on the combination of three-channel drift metrics, the drift type is determined and different adaptation paths are adopted.

[0136] When D 车辆级漂移 >T 车辆级 And D 车型级漂移 D 平台级漂移 Under normal circumstances, the information-driven greedy sampling algorithm is triggered to construct a dataset for the vehicle and perform small-sample retraining only for the vehicle-level parameter Φ1.

[0137] When D 车型级漂移 >T 车型级 And D 车辆级漂移 D 平台级漂移 Under normal circumstances, the information-driven greedy sampling algorithm is triggered, but a dataset is constructed for this region, and the vehicle-level parameter Φ2 for the same vehicle model in this region is retrained with small samples.

[0138] When D 平台级漂移 >T 平台级 And D 车辆级漂移 D 车型级漂移 Under normal circumstances, the information-driven greedy sampling algorithm is triggered, but a dataset is constructed specifically for this region, and the platform-level parameter Φ3 for this region is retrained with small samples. For the specific implementation of the greedy algorithm, please refer to Example 1.

[0139] When multiple drift metrics exceed the threshold simultaneously, a larger-scale incremental training is triggered, the residual backbone parameter θ is updated appropriately, and the relevant vehicle-level, model-level, and platform-level parameters are adjusted simultaneously. After the parameter update is completed, the updated model is deployed to the vehicle to take effect and further predict power.

[0140] like Figure 4 The diagram shown is a structural schematic of the IoT-based new energy vehicle power prediction and monitoring system provided in this application embodiment, including:

[0141] The multi-source data acquisition and preprocessing module is used to collect multi-source data from new energy vehicles through vehicle terminals, edge nodes or cloud platforms, and to perform data cleaning, outlier removal, missing value completion, time alignment and normalization fusion processing on the multi-source data to form a training sample set with complete structure, consistent time sequence and controllable quality.

[0142] The basic power calculation and feature construction module is used to calculate the basic power value of the vehicle under the current operating conditions based on the feature parameters in the training sample set, according to the vehicle energy consumption mechanism model or dynamic model, and to construct the prediction condition vector related to the residual on this basis, which is used as the input of the subsequent residual prediction model.

[0143] The residual prediction and total power generation module is used to input the prediction condition vector into the residual prediction model to obtain the residual prediction result used to correct the base power value. The base power value and the residual prediction result are fused and superimposed to generate the total predicted power of the new energy vehicle under the current working condition. At the same time, the predicted power sequence with model version and time stamp is output.

[0144] The hierarchical parameter online update module is used to obtain the actual power data of the corresponding time period during the trial operation phase of the vehicle. Before the cumulative amount of actual power data reaches the preset update condition, the calibration layer parameters are adaptively updated based on the error between the actual power data and the total predicted power to output the calibrated predicted power result, while keeping the backbone parameters of the residual prediction model unchanged.

[0145] The predicted power output and drift monitoring module is used to output calibrated predicted power results and monitoring information generated by errors, which is used for power prediction drift judgment or operation strategy adjustment.

[0146] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above functions can be divided into different functional modules to complete all or part of the functions described above.

[0147] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0148] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units, located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0149] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0150] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the solution, or all or part of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0151] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A new energy vehicle power prediction monitoring method based on an Internet of Things, characterized in that, Includes the following steps: S1: Collect multi-source data of new energy vehicles through vehicle terminals, edge nodes or cloud platforms. The multi-source data includes at least one or more of the following: vehicle operating status data, vehicle parameter data, road and environmental condition data and vehicle or area identification data. S2: The multi-source data is preprocessed by cleaning, aligning, missing data removal and normalization fusion to form a training sample set for power prediction; S3: Based on the training sample set, calculate the basic power value according to the vehicle energy consumption mechanism or dynamic mechanism to characterize the basic power component of the vehicle under the current operating conditions. S4: Input the training sample set into the residual prediction model and output the residual prediction result used to correct the base power value; S5: The base power value and the residual prediction result are fused together to generate the total predicted power of the new energy vehicle; S6: During the trial operation phase of the vehicle's online operation, the actual power data for the corresponding time period is obtained, and before the cumulative amount of the actual power data reaches the preset update condition, the calibration layer parameters are adaptively updated based on the error between the actual power data and the total predicted power to output the calibrated predicted power result, wherein the backbone parameters of the residual prediction model are kept unchanged. The calibration layer corresponding to the calibration layer parameters is a lightweight calibration structure. The lightweight calibration structure corrects the deviation of the residual prediction result or the total prediction power after fusion, and the adaptive update includes drift-based iterative update. Before adaptively updating the calibration layer parameters, a sample selection strategy is adopted to select a sample set for updating from the samples corresponding to the actual power data. The sample selection strategy is based on at least two of the following: signal completeness, signal stability, abnormal residuals, operating condition coverage and sample redundancy, so that the sample set covers different vehicle operating conditions. Signal stability refers to the degree of change of a signal within a sampling time window. For the k-th key signal, the difference is calculated within the window, and the specific formula is as follows: s k (t) represents the value of the kth key signal at the sampling point of time t, s k (t-1) represents the value of the kth key signal at the sampling point of time t-1, s k (t) represents the sampling change degree in the adjacent sampling time point, and defines a difference energy E k : E k The larger, the more unstable the signal is within the sampling time window. The signal stability is scored, noted Q stab : where ω k represents the weight, and satisfies the condition ∑ω k = 1, and K is the number of key signals that the sample should have. S7: Output the calibrated predicted power result and output the monitoring information generated by the error, which is used for power prediction drift judgment or operation strategy adjustment. 2.The new energy vehicle power prediction monitoring method based on the Internet of Things according to claim 1, characterized in that: The multi-source data includes one or more of the data collected by the vehicle terminal, edge node, or cloud platform, and at least includes one or more of the following: vehicle speed, vehicle acceleration, drive motor speed, drive motor torque, battery voltage, battery current, battery state of charge, brake pedal opening, accelerator pedal opening, gear status, ambient temperature, road slope, vehicle geographical location, and time information. The time information includes one or more of the following: sampling time window, current sample runtime, and current calendar time. 3.The new energy vehicle power prediction monitoring method based on the Internet of Things according to claim 1, characterized in that: The preprocessing includes at least two of the following: timestamp alignment, outlier removal, missing value imputation, signal denoising, unit unification, and normalization. Furthermore, the multi-source data is aggregated according to a preset time window to generate statistical features. The statistical features include at least one of the following: extreme value features, fluctuation features, rate of change features, or stability features. 4.The new energy vehicle power prediction monitoring method based on the Internet of Things according to claim 1, characterized in that: When calculating the base power value according to the vehicle energy consumption mechanism or dynamic mechanism, at least two of the following should be considered: vehicle rolling resistance, air resistance, slope resistance, acceleration resistance, and transmission efficiency. The calculation should be based on at least one or more of the following: vehicle curb weight, tire rolling resistance coefficient, frontal area, air resistance coefficient, road slope information, or transmission efficiency parameters. 5.The new energy vehicle power prediction monitoring method based on the Internet of Things according to claim 1, characterized in that: The residual prediction model is a time series prediction model based on machine learning. The time series prediction model uses time series features formed by multi-source data within a preset historical time window to predict residuals. The time series prediction model includes at least one or more of the following: recurrent neural network model, lightweight deep network model, gradient boosting regression model, or ensemble regression model. 6.The new energy vehicle power prediction monitoring method based on the Internet of Things according to claim 1, characterized in that: The monitoring information includes error statistics and input data distribution deviation indicators calculated based on a time window. The error statistics include at least one of average deviation, bias degree, or error fluctuation degree. The input data distribution deviation indicator is used to characterize the distribution change of the current input data relative to the historical benchmark data, and to determine drift based on the monitoring information.

7. The IoT-based power prediction and monitoring method for new energy vehicles as described in claim 6, characterized in that: When the error statistics index or the input data distribution deviation index meets the preset triggering conditions within multiple consecutive monitoring time windows, the model update process is executed. The model update process includes uploading the multi-source data and its corresponding actual power data during the monitoring period to the cloud or edge node, retraining or incrementally updating the residual prediction model or calibration layer parameters, generating a model version and distributing it to the vehicle terminal or edge node for deployment. The model update process further includes at least one of version consistency verification and rollback strategies.

8. The system for applying the new energy vehicle power prediction monitoring method based on the Internet of Things according to any one of claims 1-7, characterized in that, include: The module includes a multi-source data acquisition and preprocessing module, a basic power calculation and feature construction module, a residual prediction and total power generation module, a hierarchical parameter online update module, and a predicted power output and drift monitoring module. The multi-source data acquisition and preprocessing module is used to collect multi-source data of new energy vehicles through vehicle terminals, edge nodes or cloud platforms, and perform data cleaning, outlier removal, missing value completion, time alignment and normalization fusion processing on the multi-source data to form a training sample set with complete structure, consistent time sequence and controllable quality. The basic power calculation and feature construction module is used to calculate the basic power value of the vehicle under the current operating conditions based on the feature parameters in the training sample set, according to the vehicle energy consumption mechanism model or dynamic model, and to construct a prediction condition vector related to the residual on this basis, which is used as the input of the subsequent residual prediction model. The residual prediction and total power generation module is used to input the prediction condition vector into the residual prediction model to obtain the residual prediction result used to correct the base power value, fuse and superimpose the base power value and the residual prediction result to generate the total predicted power of the new energy vehicle under the current working condition, and output the predicted power sequence with model version and time stamp. The hierarchical parameter online update module is used to acquire the actual power data for the corresponding time period during the trial operation phase of the vehicle's online operation, and before the cumulative amount of the actual power data reaches the preset update condition, to adaptively update the calibration layer parameters based on the error between the actual power data and the total predicted power, so as to output the calibrated predicted power result, wherein the backbone parameters of the residual prediction model are kept unchanged. The predicted power output and drift monitoring module is used to output calibrated predicted power results and monitoring information generated by the error, which is used for power prediction drift judgment or operation strategy adjustment.

Images