An electric vehicle charging curve prediction method and system based on a PI-beta-VAE model, a storage medium, and an electronic device
Patent Information
- Application Number
- CN202611072021.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-20
- Publication Date
- 2026-10-23
Smart Images
Figure CN122965743A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electric vehicle charging management and intelligent data analysis technology, specifically a method based on... The model includes methods, systems, storage media, and electronic devices for predicting electric vehicle charging curves. Background Technology
[0002] With the rapid development of my country's new energy vehicle industry, the number of electric vehicles on the road continues to rise, and the supporting public DC fast charging infrastructure is being deployed on a large scale across the country. Accurate charging curve prediction is of vital importance for improving the charging service experience, optimizing grid load dispatching, and increasing the utilization rate of charging facilities.
[0003] A charging curve typically refers to the sequential relationship between charging power and the battery's state of charge (SoC). It not only determines the duration of the charging process but also directly impacts charging station operators' predictions of future power demand. In actual operation, if the complete power-SoC evolution trajectory can be accurately deduced from a limited set of observational data points at the initial stage of a vehicle's charging, operators can provide drivers with precise charging completion time estimates. This helps logistics fleet managers make vehicle dispatching decisions and assists the power grid dispatch center in developing minute-level power balancing plans and demand response strategies. Furthermore, accurate charging curve prediction is also a crucial basis for allocating charging resources within charging stations, estimating queuing times, and dynamically adjusting charging service fees.
[0004] However, in real-world applications, charging curves are highly nonlinear and random due to the coupled influence of various factors, including vehicle model, battery capacity, ambient temperature, battery health, charging pile rated power, and user charging habits. Traditional physics-based methods require detailed battery parameters and thermodynamic models, but in public charging scenarios, the variety of vehicle models and the inability to obtain specific battery parameters make physics models unsuitable. Existing purely data-driven methods mostly rely on complete charging history data or require fixed-length feature vectors, failing to flexibly address the situation where data is gradually generated during actual charging and early data is scarce. More critically, when faced with extremely sparse initial data, purely data-driven deep learning methods may generate curves that violate electrochemical or physical principles due to the randomness of neural networks. For example, power may suddenly rebound during the constant voltage phase at the end of charging, or the curve may exhibit physically impossible, sharp sawtooth fluctuations. While a few deep learning-based methods can handle time-series predictions to some extent, they often lack the ability to quantify prediction uncertainty and do not embed the macroscopic physical laws of battery charging and discharging as prior knowledge into the model training process, limiting the value of prediction results in risk-sensitive applications.
[0005] Therefore, there is an urgent need for a model-based electric vehicle charging curve prediction method, system, storage medium, and electronic device that can generate a complete charging curve prediction that conforms to physical laws based on only a small amount of partially observed charging data, and can provide a method for estimating prediction uncertainty, in order to solve the problems existing in the above-mentioned prior art. Summary of the Invention
[0006] The purpose of this invention is to propose a model-based method, system, storage medium, and electronic device for predicting electric vehicle charging curves, thereby achieving strong coupling between data fitting and physical laws, so as to reconstruct a complete charging curve that conforms to macroscopic physical laws with only a very small number of observation points.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] In a first aspect, the present invention provides a method based on The method for predicting electric vehicle charging curves using the model includes the following steps:
[0009] Build The model is trained using historical charging data to obtain the trained result. Model;
[0010] The The model includes an encoder network, a decoder network, an invertible instance normalization module, and a physical information joint loss function. The joint loss function used during training consists of the sum of a reconstruction error term, a weighted Kullback-Leibler divergence loss term, and a physical constraint loss term, and the strength of the weighting is dynamically adjusted through an annealing mechanism.
[0011] When the electric vehicle charging process reaches a certain point, the system collects some of the charging curve data accumulated since the start of charging in real time, and obtains static covariate data related to this charging session.
[0012] The real-time collected partial charging curve data and static covariate data are input into the deployed, trained system. The model triggers real-time prediction of the complete charging curve in the future;
[0013] use The model's encoder network compresses the input data into a latent space, generating latent variables; the latent space is explicitly decoupled into a physical manifold subspace and a behavioral manifold subspace.
[0014] use The model's decoder network decodes and reconstructs the latent variables to generate a complete charging curve prediction result;
[0015] The predicted charging time is calculated based on the prediction results of the complete charging curve.
[0016] As a further improvement, the partial charging curve data is a power-state-of-charge sequence consisting of several data points, and the static covariate data includes at least the initial state of charge, the rated power of the charging pile, the connector type, the ambient temperature, and the online estimated rated battery capacity.
[0017] As a further improvement, both the encoder network and the decoder network include a long short-term memory network layer; the encoder network outputs the mean vector and log-variance vector of the latent variable distribution, and obtains the latent variables from the distribution through reparameterization.
[0018] As a further improvement, the latent space dimension K of the latent space is explicitly split into two parts:
[0019] Physical manifold subspace From the initial state of charge Ambient temperature and battery capacity Physical covariates are characterized by direct linear mapping or strong constraint guidance through fully connected layers, representing the inherent thermodynamic and electrochemical background states of the battery.
[0020] and behavioral manifold subspace The latent variables are sampled from an adaptive Gaussian distribution using reparameterization techniques to characterize a user's personalized charging habits or random perturbations during a single session; finally, complete latent variables are synthesized. It is sent to the decoder.
[0021] As a further improvement, the reconstruction error term is calculated based on the mean square error and the influence of the padding position is ignored using the mask vector; the weighted Kullback-Leibler divergence loss term is used to measure the degree of difference between the encoder output distribution and the standard normal distribution; the physical constraint loss term is composed of three penalty operators: power monotonically decay constraint, power first derivative continuity constraint, and energy conservation boundary constraint.
[0022] As a further improvement, the annealing mechanism sets the weighting strength of the weighted Kullback-Leibler divergence loss term close to zero in the early stage of training, so that the model prioritizes learning data reconstruction; as the training cycle increases, the weighting strength gradually increases to strengthen the regularization constraint on the latent space.
[0023] As a further improvement, the reversible instance normalization module performs instance normalization on the power sequence of each charging curve before the data enters the encoder, and performs inverse normalization after the decoder outputs the reconstruction result to restore the original power scale.
[0024] As a further improvement, the calculation of the predicted charging time based on the prediction results of the complete charging curve is achieved through integral calculation:
[0025] ;
[0026] in, For the predicted charging session duration, and These are the current state of charge and the end of the state of charge, respectively. For the estimated electric vehicle battery capacity, Let d be the predicted charging power at the state of charge (SOC), d be the derivative, dSOC be the derivative of the state of charge (SOC), and SOC be the state of charge.
[0027] Secondly, the present invention also provides a method based on A model-based electric vehicle charging curve prediction system, characterized in that the system is used to perform a prediction based on... A method for predicting electric vehicle charging curves using a model;
[0028] include:
[0029] The model building and training module is used to build and train models. Model;
[0030] The data acquisition module is used to collect accumulated partial charging curve data and static covariate data in real time during the charging process;
[0031] The model invocation module is used to input the real-time acquired data into the deployed model and initiate the prediction of the future charging curve.
[0032] The encoding module is used to drive the encoder network to generate latent variables;
[0033] The decoding prediction module is used to drive the decoder network to reconstruct the complete charging curve;
[0034] The duration calculation module is used to calculate the predicted charging duration based on the reconstruction results.
[0035] Thirdly, the present invention also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform tasks based on... A method for predicting electric vehicle charging curves using a model.
[0036] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform tasks based on... A method for predicting electric vehicle charging curves using a model.
[0037] Compared with the prior art, the beneficial effects of the present invention are:
[0038] 1. Overcoming the generalization blind spot under "extreme sparsity": When the car owner has just plugged in the charger and only 1 to 2 data points are generated, the curves generated by the pure data-driven model are easily distorted due to the lack of input information. This invention introduces physical information loss function constraints and decouples them from the potential space of the physical manifold. Physical priors such as the electrochemical degradation trend of battery charging and discharging and the continuity of the constant current-constant voltage switching point are written into the loss function as regularization terms. This allows physical laws to play the role of "mentor" during the training process, forcibly confining the prediction within a reasonable physical manifold, achieving high-precision prediction at the first minute level.
[0039] 2. Eliminating anomalous distortions and ensuring reliable output: The monotonicity penalty operator in the physical constraint loss term ensures that the power decreases monotonically in the later stages of fast charging, the smoothness penalty operator eliminates abrupt and non-differentiable abrupt changes in the curve, and the energy conservation constraint ensures that the predicted power does not exceed the physical upper limit allowed by the current ambient temperature and the rated power of the charging pile. These constraints eliminate the occasional "outliers" and unrealistic power fluctuations generated by the neural network, making the predicted curve perfectly fit the mainstream charging process specifications such as constant current-constant voltage, and more suitable for direct connection to the grid dispatch system, which is extremely sensitive to risks.
[0040] 3. Significantly enhanced interpretability: Through explicit "physical-behavioral" decoupling of the latent space, the physical manifold subspace is transformed from... Ambient temperature and battery capacity By directly mapping physical covariates, the behavioral manifold subspace is characterized by reparameterized sampling to represent user random perturbations. This design clearly shows engineers which model outputs are determined by the objective physical state of the battery and which are determined by user random behavior, completely shedding the academic label of traditional large models as "pure black boxes."
[0041] 4. Strong generalization to vehicle models and operating conditions: By introducing a reversible instance normalization module at the model front end, The model can automatically eliminate scale differences in charging power amplitude between vehicles with different battery capacities, focusing its learning on the shape characteristics of the charging curve. This allows a single model to be applied to various electric vehicles of different brands, capacities, and charging rates, without the need for separate modeling or maintaining complex vehicle parameter databases for each model.
[0042] 5. Robust latent space structure: The model training process adopts a joint loss function combined with annealing strategy. In the early stage of training, the reconstruction accuracy is prioritized, and the regularization constraint of the latent space is gradually enhanced as training progresses. This effectively avoids the posterior collapse problem, so that the learned latent space has better continuity and structure, laying the foundation for anomaly detection and high-quality generative prediction based on reconstruction error.
[0043] 6. Dynamic Real-Time Update Capability: As the charging process progresses, new observational data is continuously generated. This invention supports re-inputting the updated partial charging curve into the model, quickly refreshing the prediction results of the complete charging curve and the corresponding charging time estimate. This real-time update mechanism allows the prediction accuracy to continuously improve with the increase of data, providing users with gradually converging and more accurate remaining time prompts.
[0044] 7. Low deployment cost and strong applicability: The input data required for this invention can be directly obtained at the charging pile or obtained through simple estimation, without relying on proprietary protocol data provided by the vehicle battery management system. After the model is trained, it can be deployed on a cloud server or edge computing device to provide real-time prediction services for large-scale charging networks, which has extremely high engineering practical value. Attached Figure Description
[0045] Figure 1 The first embodiment of the present invention is based on Flowchart of the electric vehicle charging curve prediction method in the model;
[0046] Figure 2 Provided for Embodiment 1 of the present invention A schematic diagram of the model's network structure and training process;
[0047] Figure 3 Provided for Embodiment 1 of the present invention A schematic diagram of the data flow during the real-time prediction phase of the model. Detailed Implementation
[0048] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:
[0049] Example 1
[0050] This embodiment provides a method based on A method for predicting electric vehicle charging curves using a model. For example... Figure 1 As shown, this method includes two stages: model building and training, and real-time prediction. The real-time prediction stage mainly includes the following steps:
[0051] S1. When the electric vehicle charging process reaches a certain point, collect in real time a portion of the charging curve data accumulated since the start of charging, and obtain static covariate data related to this charging session.
[0052] S2. Input the real-time collected partial charging curve data and static covariate data into the deployed, pre-trained system. The model triggers a real-time prediction process for the complete charging curve in the future.
[0053] S3, using the above The model's encoder network compresses the input data into a latent space, generating latent variables; the latent space is explicitly decoupled into a physical manifold subspace and a behavioral manifold subspace.
[0054] S4, using the above The model's decoder network decodes and reconstructs the latent variables to generate a complete charging curve prediction result;
[0055] S5. Calculate the predicted charging time based on the prediction results of the complete charging curve.
[0056] The following section will first provide a detailed explanation of the model building and training phases, and then elaborate on each step of the real-time prediction phase.
[0057] like Figure 2 As shown, this involves the first stage: model building and training.
[0058] 1. Core Innovation Architecture and Principles
[0059] The core idea is to strongly couple "data fitting" with "physical laws". We no longer treat neural networks as complete black boxes, but instead incorporate physical priors such as the electrochemical decay trend of power P with battery state of charge (SOC) during DC fast charging of electric vehicles, and the continuity of the constant current-constant voltage switching point, as regularization terms into the loss function. At the same time, we perform explicit decoupling design in the latent space dimension, separating latent variables into a physical manifold subspace directly constrained by physical covariates and a behavioral manifold subspace representing random perturbations. This ensures that even with only 1 to 2 observation points, the reconstructed complete charging curve absolutely conforms to the macroscopic physical laws of battery charging.
[0060] 2. Model Network Structure Construction
[0061] This embodiment constructs The model consists of five parts: an encoder network, a latent space sampling layer, a decoder network, an invertible instance normalization module, and a joint loss function for physical information.
[0062] The encoder network consists of three stacked Long Short-Term Memory (LSTM) layers, with 128 hidden units in each layer. After the input sequence is encoded by LSTM, the hidden state at the last time step is taken as the global feature representation of the sequence. This hidden state is mapped to the mean vector of the latent distribution through two parallel fully connected layers. Sum of logarithmic variance vector Therefore, a multidimensional Gaussian distribution is defined as the approximate posterior distribution:
[0063] ;
[0064] in, This is the approximate posterior distribution of the encoder network output. It is a diagonal covariance matrix. Let be the variance vector of the latent space. These are the trainable parameters of the encoder network. To input charging curve data, For static covariates, Let N be the identity matrix, and let N be the symbol for the normal distribution.
[0065] In order to obtain specific latent variables from the distribution of encoder output Sampling is performed using reparameterization techniques:
[0066] ;
[0067] in, To obtain from the standard normal distribution Random noise vector sampled in the middle, , This indicates element-wise multiplication.
[0068] The core innovation of this embodiment lies in the explicit decoupling design of the latent space, which explicitly splits the latent space dimension K into two parts:
[0069] Physical manifold subspace From the initial state of charge Ambient temperature and battery capacity Physical covariates are obtained through direct linear mapping or strong constraint guidance via fully connected layers, characterizing the inherent thermodynamic and electrochemical background states of the battery. Specifically, let the physical covariate vector be... ,but:
[0070] ;
[0071] in, For physical manifold subspaces, The rated power of the charging pile Let be the learnable weight matrix from the physical covariates to the physical manifold subspace. The corresponding learnable bias vectors, together, constitute the linear mapping parameters of the fully connected layer.
[0072] Behavioral manifold subspace Sampled from an adaptive Gaussian distribution using reparameterization techniques, it characterizes a specific user's personalized charging habits or random perturbations during a single session.
[0073] ;
[0074] in, and These are the mean vector and logarithmic standard deviation vector output by the encoder. It is a random noise vector sampled from a standard normal distribution.
[0075] Finally, the complete latent variables are synthesized:
[0076] ;
[0077] Experiments have verified that setting the dimension of the physical manifold subspace to [value] is [effective]. The dimension of the behavior manifold subspace is set to Total potential space dimension This effectively preserves the core characteristics of the charging behavior.
[0078] The decoder network also consists of three stacked long short-term memory (LSTM) layers, with 128 hidden units in each layer. The initial hidden state of the decoder is determined by latent variables. Initialized after linear transformation. The decoder generates the reconstructed sequence step by step in an autoregressive manner, and finally outputs the complete reconstructed charging curve. .
[0079] The reversible instance normalization module is located at both the encoder input and decoder output. During the input phase, this module independently calculates the mean *m* and standard deviation *s* of the power sequence for each charging curve, and uses these two statistics to perform instance normalization on the sequence.
[0080] ;
[0081] in This is the original power value. This is the normalized power value. This is a very small constant used to prevent division by zero errors. After normalization, the charging curves corresponding to different vehicle models and battery capacities are mapped to a uniform scale, allowing the encoder to focus on learning the curve shape features while ignoring absolute numerical differences. After the decoder outputs the reconstructed sequence, this module uses the previously saved mean m and standard deviation s to inversely normalize the output, restoring the original power dimensions.
[0082] ;
[0083] 3. Training Data Preparation
[0084] Large-scale completed DC fast charging session records were collected from the historical operation database of the charging network. Each record contains a complete power-state-of-charge curve and the corresponding static covariates. The collected data was cleaned to remove abnormal sessions with missing key fields or communication interruptions. Each charging curve was resampled at fixed state-of-charge intervals and converted into a fixed-length sequence; for sections with insufficient length, mask values were used to pad and corresponding Boolean mask vectors were generated. , used to mark the location of the actual data.
[0085] 4. Definition of Joint Loss Function for Physical Information
[0086] The training objective of the model is to minimize the joint loss function of physical information, which is driven by three parts:
[0087] ;
[0088] in, For the joint loss function of physical information, To reconstruct the loss term, For KL divergence loss term, For physical constraint loss terms, To control the hyperparameters of the latent space regularization strength, The hyperparameters used to control the strength of physical constraints.
[0089] (1) Data-driven items
[0090] Reconstructing the loss term Used to measure decoder output With the original input The degree of difference between them. This embodiment uses mean squared error as a metric, and through a mask vector. The interference of the shielding filling location on loss calculation:
[0091] ;
[0092] in, For sequence index, The first of the mask vectors Each element.
[0093] KL (Kullback-Leibler) divergence loss term This loss term, used to measure the difference between the approximate posterior distribution of the encoder output and the standard normal prior distribution, has an analytical form and can be calculated directly.
[0094] ;
[0095] In the above formula, K represents the dimension of the latent space. and The encoder at the 1st The mean and standard deviation of the output.
[0096] (2) Physical constraints ( (Specific structure)
[0097] This invention utilizes the classical physical properties of battery charging and discharging to construct the following three penalty operators:
[0098] Operator A: Monotonic power decay and boundary constraints (polarization penalty)
[0099] Based on the battery's electrochemical characteristics, in the middle and later stages of fast charging (usually...) During the constant voltage phase (at which the polarization voltage increases), the charging power must monotonically decrease. Its mathematical expression is:
[0100] ;
[0101] in, For power single attenuation constrained loss, For the first One charge state sampling point, In the state of charge Predicted charging power at the location.
[0102] Physical meaning: If the power derivative predicted by the model at a high SOC level is greater than 0 (i.e. the power increases instead of decreasing), it will be subject to a huge numerical penalty.
[0103] Operator B: Power first derivative continuity constraint (CV-CC switching smoothing penalty)
[0104] In actual charging, the onboard BMA adjusts the power continuously, and the prediction curve cannot exhibit abrupt breaks or drastic changes that prevent it from becoming differentiable. Its mathematical expression is:
[0105] ;
[0106] in, The power first derivative continuity constraint loss.
[0107] Physical meaning: Penalize local "jagged" or abrupt changes in the slope of the prediction curve to ensure that the generated curve has a smooth physical texture.
[0108] Operator C: Capacity-Integral Boundary Constraints Based on Energy Conservation
[0109] The area enclosed by the complete curve (integrated over time) represents the electrical energy charged, which is related to the battery capacity. Furthermore, the voltage sequence must satisfy the law of conservation of energy. The predicted power must absolutely not exceed the physical upper limit allowed by the current ambient temperature and the rated power of the charging pile. Its mathematical expression is:
[0110] ;
[0111] in, Loss due to energy conservation boundary constraints. This represents the maximum allowable charging time under current environmental conditions.
[0112] The comprehensive physical constraint loss term is:
[0113] ;
[0114] in , , These are the weight coefficients for the three penalty operators.
[0115] 5. Beta Annealing Training Strategy
[0116] In the initial training phase, the weights of both the encoder and decoder are randomly initialized, and the model's reconstruction ability is still weak. If a large penalty is imposed on the KL divergence term at this time, the encoder will tend to directly output latent variables that are consistent with the standard normal distribution but are unrelated to the input content, thus causing the model to fail to learn meaningful latent representations. This phenomenon is called "posterior collapse".
[0117] To effectively avoid the posterior collapse problem, this embodiment adopts an annealing strategy, dynamically adjusting the parameters during training. The value of . Specifically, With training cycle The increment is gradually increased according to the following logistic growth function:
[0118] ;
[0119] in, For the current training cycle, A constant used to control the steepness of the annealing curve, for The training cycle in which the maximum value is reached is half of the maximum value. for The target maximum value. At the start of training. Approaching zero, the model prioritizes learning to accurately reconstruct the data; as training progresses, As the value increases, the regularization constraint on the latent space gradually strengthens. In this embodiment, the hyperparameter is set to... , , .
[0120] 6. Model Optimization and Saving
[0121] Training employs a self-supervised learning strategy: for each complete charging curve, a random subset of continuous prefix data is selected as the known input, and the remaining portion is used as the target to be reconstructed. The AdamW optimizer is used, with an initial learning rate set to... The batch size is 512. The model converges after approximately 250 training epochs. After training, the encoder network parameters φ and decoder network parameters θ are saved to obtain parameters that can be used for real-time prediction. Model.
[0122] II. Real-time Prediction Phase
[0123] Step S1: When the electric vehicle charging process reaches a certain point, collect some of the charging curve data accumulated since the start of charging in real time, and obtain the static covariate data related to this charging session.
[0124] In this embodiment, after the electric vehicle connects to the DC fast charging pile and begins charging, the charging pile controller records charging power and battery state of charge data at fixed time intervals or with a fixed change in state of charge as the step size. At a certain point in the charging process, the system retrieves all data points accumulated since the start of charging in real time. These data points constitute the partial charging curve data. This partial charging curve data is represented as a sequence of two-dimensional coordinate points, where the horizontal axis represents the state of charge and the vertical axis represents the charging power.
[0125] Simultaneously, the system acquires static covariate data relevant to this charging session. These static covariates remain constant or change extremely slowly throughout the charging process. Specifically, they include: initial state of charge. This refers to the battery's state of charge (SOC) value when the vehicle is connected to the charging station; the rated power of the charging station. This refers to the maximum output power currently supported by the charging station; the connector type, such as a national standard DC interface or a combined charging system interface; and the ambient temperature. The estimated electric vehicle battery capacity can be obtained through the temperature sensor built into the charging station or by connecting to an external meteorological data interface. It can be estimated online based on the delivered energy and the change in state of charge.
[0126] Step S2: Input the real-time collected partial charging curve data and static covariate data into the deployed, pre-trained... The model triggers a real-time prediction process for the complete charging curve in the future.
[0127] The embodiment used The model has completed all training work in the aforementioned model building and training phases and has been deployed on a cloud server or charging station edge computing device. After step S1 completes the data acquisition for the current moment, the system immediately inputs the partial charging curve data and static covariate data into the deployed model, triggering the prediction of future charging curves.
[0128] Before formally inputting the model, the system first preprocesses the input sequence: the power-state sequence is resampled to a uniform grid of fixed length using interpolation methods, and unobserved future portions are marked as unknown using mask vectors. Static covariate data is encoded into context vectors that match the dimensions of the time series input, guiding the information processing of the encoder and decoder.
[0129] Figure 3 Showing The model's structure and data flow during the prediction phase. Partial charging curve data is first processed by a reversible instance normalization module. This module calculates the mean and standard deviation of the current input sequence and normalizes the power values to eliminate scale differences in power amplitude across different vehicle models, allowing the model to focus on learning the shape characteristics of the curve.
[0130] Step S3: Utilize the above The model's encoder network compresses the input data into the latent space, generating latent variables.
[0131] Normalized partial charging curve data, along with the static covariate context vector, are input into the encoder network. In this embodiment, the encoder network is composed of multiple stacked long short-term memory networks, which can effectively capture the temporal dependency of charging power as a function of state of charge. Even if there are a large number of unknown positions marked by masks in the input sequence, the encoder can still extract meaningful feature representations based on the observed portions.
[0132] The encoder's final hidden state is mapped to the mean vector of the latent variable distribution through two independent fully connected layers. Sum of logarithmic variance vector These two vectors together define a multidimensional Gaussian distribution, representing the model's inference of the corresponding location in the latent space for the current portion of the observed data.
[0133] The core innovation of this step lies in the explicit decoupling of the latent space. (Physical covariate vector) Physical manifold subspaces are generated directly through independent linear mapping layers. This subspace is unaffected by the randomness of reparameterized sampling, ensuring that physical prior information is injected into the latent representation in a deterministic manner. Behavioral manifold subspace The latent variables are then sampled from the encoder output distribution using standard reparameterization techniques to characterize the random factors specific to the current session. The final synthesized complete latent variables... It is sent to the decoder.
[0134] Step S4: Utilize the above The model's decoder network decodes and reconstructs the latent variables to generate a complete charging curve prediction result.
[0135] The obtained latent variable z, along with the static covariate context vector, is input into the decoder network. The decoder also consists of a multi-layer long short-term memory network, and its initial hidden state is initialized by linearly transforming the latent variable. The decoder progressively generates the predicted charging power value corresponding to each state of charge point in an autoregressive manner until a complete charging curve covering the entire range of states of charge is generated.
[0136] The power sequence output by the decoder is denormalized by the reversible instance normalization module to restore it to the original power scale. The denormalization process uses the mean and standard deviation of the input sequence saved in step S2 to ensure that the output predicted value is consistent with the actual physical dimensions.
[0137] If it is necessary to quantify the uncertainty of the prediction, different latent variables can be sampled multiple times from the distribution of the encoder output. (Only for) Some samples were taken multiple times. (Keep the parameters constant), and generate multiple possible charging curves using a decoder. The statistical distribution of these curves reflects the model's prediction confidence interval for the complete charging curve under the current partial observations.
[0138] Step S5: Calculate the predicted charging time based on the predicted results of the complete charging curve.
[0139] Based on the complete charging curve generated in step S4, the estimated time required to charge from the current state of charge to the target state of charge can be obtained through integration calculation:
[0140] ;
[0141] in, For the predicted charging session duration, and These are the current state of charge and the end of the state of charge, respectively. For the estimated electric vehicle battery capacity, Let d be the predicted charging power at the state of charge (SOC), d be the derivative, dSOC be the derivative of the state of charge (SOC), and SOC be the state of charge.
[0142] As the charging process progresses, new power-state-of-charge data points are continuously generated. The system can then re-input the updated partial charging curve. The model repeats steps S2 to S5 to update the prediction results in real time. Typically, as the amount of observed data increases, the confidence interval of the prediction gradually narrows, and the predicted charging time becomes more accurate. The model updates the complete curve and the accurate end-of-charge time (EOTC) online every minute.
[0143] Example 2
[0144] This embodiment provides a method based on The model's electric vehicle charging curve prediction system corresponds one-to-one with the steps described in Example 1. The system includes:
[0145] The model building and training module is used to build and train models. The model, specific network structure, and training process are as described in Example 1;
[0146] The data acquisition module is used to collect accumulated partial charging curve data and static covariate data in real time during the electric vehicle charging process;
[0147] The model invocation module is used to input the real-time collected data into the deployed, trained model to trigger the prediction of the future complete charging curve.
[0148] An encoding module is used to generate latent variables using the encoder network of the model;
[0149] The decoding prediction module is used to reconstruct the complete charging curve using the decoder network of the model;
[0150] The duration calculation module is used to calculate the predicted charging duration based on the complete charging curve.
[0151] It should be noted that the above modules correspond to the steps described in Embodiment 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.
[0152] In further embodiments, the following is also provided:
[0153] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.
[0154] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.
[0155] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, memory may also store information about the device type.
[0156] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.
[0157] The method in Example 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.
[0158] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.
[0159] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.
Claims
1. A method based on The electric vehicle charging curve prediction method based on the model is characterized by, Includes the following steps: Build The model is trained using historical charging data to obtain the trained result. Model; The The model includes an encoder network, a decoder network, an invertible instance normalization module, and a physical information joint loss function. The joint loss function used during training consists of the sum of a reconstruction error term, a weighted Kullback-Leibler divergence loss term, and a physical constraint loss term, and the strength of the weighting is dynamically adjusted through an annealing mechanism. When the electric vehicle charging process reaches a certain point, the system collects some of the charging curve data accumulated since the start of charging in real time, and obtains static covariate data related to this charging session. The real-time collected partial charging curve data and static covariate data are input into the deployed, trained system. The model triggers real-time prediction of the complete charging curve in the future; use The model's encoder network compresses the input data into the latent space, generating latent variables; The potential space is explicitly decoupled into a physical manifold subspace and a behavioral manifold space; use The model's decoder network decodes and reconstructs the latent variables to generate a complete charging curve prediction result; The predicted charging time is calculated based on the prediction results of the complete charging curve. The predicted charging time, calculated based on the complete charging curve prediction results, is achieved through integral calculation. ； in, For the predicted charging session duration, and These are the current state of charge and the end of the state of charge, respectively. For the estimated electric vehicle battery capacity, This is the change in state of charge. Let d be the predicted charging power at the state of charge (SOC), d be the derivative, dSOC be the derivative of the state of charge (SOC), and SOC be the state of charge.
2. A method based on claim 1 The electric vehicle charging curve prediction method based on the model is characterized by, The partial charging curve data is a power-state-of-charge sequence consisting of several data points. The static covariate data includes at least the initial state of charge, the rated power of the charging pile, the connector type, the ambient temperature, and the online estimated rated battery capacity.
3. A method based on claim 1 The electric vehicle charging curve prediction method based on the model is characterized by, Both the encoder network and the decoder network contain long short-term memory network layers; the encoder network outputs the mean vector and log-variance vector of the latent variable distribution, and obtains the latent variables from the distribution through reparameterization.
4. A method based on claim 1 The electric vehicle charging curve prediction method based on the model is characterized by, The potential space dimension K of the potential space is explicitly split into two parts: Physical manifold subspace From the initial state of charge Ambient temperature and battery capacity Physical covariates are characterized by direct linear mapping or strong constraint guidance through fully connected layers, representing the inherent thermodynamic and electrochemical background states of the battery. and behavioral manifold subspace The latent variables are sampled from an adaptive Gaussian distribution using reparameterization techniques to characterize a user's personalized charging habits or random perturbations during a single session; finally, complete latent variables are synthesized. It is sent to the decoder.
5. A method based on claim 1 The electric vehicle charging curve prediction method based on the model is characterized by, The reconstruction error term is calculated based on the mean square error and uses the mask vector to ignore the influence of the padding position; the weighted Kullback-Leibler divergence loss term is used to measure the degree of difference between the encoder output distribution and the standard normal distribution; the physical constraint loss term is composed of three penalty operators: power monotonically decay constraint, power first derivative continuity constraint, and energy conservation boundary constraint.
6. A method based on claim 1 The electric vehicle charging curve prediction method based on the model is characterized by, The annealing mechanism sets the weighting strength of the weighted Kullback-Leibler divergence loss term close to zero in the early stage of training, so that the model prioritizes learning data reconstruction; as the training cycle increases, the weighting strength gradually increases to strengthen the regularization constraint on the latent space.
7. A method based on claim 1 The electric vehicle charging curve prediction method based on the model is characterized by, The reversible instance normalization module performs instance normalization on the power sequence of each charging curve before the data enters the encoder, and performs inverse normalization after the decoder outputs the reconstruction result to restore the original power scale.
8. A method based on The electric vehicle charging curve prediction system based on the model is characterized by, The system is used to perform the operation based on any one of claims 1-8. A method for predicting electric vehicle charging curves using a model; include: The model building and training module is used to build and train models. Model; The data acquisition module is used to collect accumulated partial charging curve data and static covariate data in real time during the charging process; The model invocation module is used to input the real-time acquired data into the deployed model and initiate the prediction of the future charging curve. The encoding module is used to drive the encoder network to generate latent variables; The decoding prediction module is used to drive the decoder network to reconstruct the complete charging curve; The duration calculation module is used to calculate the predicted charging duration based on the reconstruction results.
9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the tasks described in any one of claims 1-7 based on... A method for predicting electric vehicle charging curves using a model.
10. An electronic device, characterized in that, Includes a memory, a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the functions described in any one of claims 1-7 based on... A method for predicting electric vehicle charging curves using a model.