Electric vehicle charging load generation method and system

Abstract

The invention discloses an electric vehicle charging load generation method and system. The method comprises the steps that a one-dimensional time sequence signal is converted into a heterogeneous feature token containing local, global and guiding information through mixed token processing; constructing a decomposition type diffusion model based on the token, and directly generating three core components of trend, seasonality and residual error of the time sequence; optimizing a model training process by adopting a composite loss function and a warm start mechanism so as to solve the cold start problem of model training; and performing high-quality fusion reconstruction on each component through a weighted overlap-add method, and outputting a final charging load curve. According to the method, highly real and diversified charging load data can be generated, and powerful data support is provided for power grid planning and dispatching in a structure interpretable mode.

Description

Technical Field

[0001] This invention relates to a method for generating charging load for electric vehicles, belonging to the fields of electric vehicle charging management and smart grid technology. Background Technology

[0002] With increasing global emphasis on low-carbon initiatives and technological breakthroughs in batteries and electronic controls, the penetration rate of electric vehicles (EVs) in global transportation systems continues to rise. This trend poses serious challenges to charging infrastructure, particularly the planning and stable operation of power systems. The charging behavior of large-scale EVs exhibits spatiotemporal uncertainty, and the load resulting from high-concurrency, random charging demands becomes a key factor affecting grid security. Therefore, accurately and efficiently predicting and simulating EV charging loads is a technological prerequisite for addressing these challenges, unlocking flexibility potential, and ensuring the safe and economical operation of the power grid. However, in practical applications, obtaining large-scale, high-precision real-world charging data faces multiple challenges, including high costs, long collection cycles, inconsistent data quality, and privacy concerns.

[0003] To address the issue of data scarcity, existing technologies employ various data generation methods, such as Markov chains or the ARIMA time series analysis model. However, these models make relatively strict assumptions and struggle to capture the highly nonlinear and non-stationary nature of charging behavior. Consequently, the generated load data often deviates significantly from real-world scenarios. Furthermore, some simple resampling methods generate samples lacking diversity, failing to cover charging peaks during extreme scenarios such as holidays and severe weather, leading to insufficient accuracy in grid risk assessments based on this data.

[0004] While deep learning models have shown great potential in data generation, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), their black-box nature limits their application in power systems. Although the curves they generate may resemble real data in shape, their underlying structure is uninterpretable, including base load trends and periodic patterns in user behavior. This significantly restricts refined analysis and decision-making in load decomposition and demand response potential assessment. Furthermore, existing models typically use raw time series data directly as input, failing to fully exploit multi-scale patterns and global contextual information within the sequences, thus limiting their learning capabilities. For deep models that need to perform complex tasks such as component decomposition, the cold start problem in the early stages of training is particularly prominent. The lack of effective prior knowledge guidance leads to slow convergence and a tendency to get trapped in local optima.

[0005] In summary, a novel method for generating electric vehicle charging loads is needed. This method should not only generate high-fidelity and diverse data, but also possess good interpretability and address the shortcomings of existing deep learning models in training and representation. Summary of the Invention

[0006] The technical problem to be solved by this invention is to overcome the shortcomings of existing deep learning models in the process of generating electric vehicle charging load, including insufficient authenticity of generated data, lack of diversity, single representation of input information, and difficulty in cold start of model training.

[0007] To address the aforementioned technical problems, this invention provides a method for generating charging load for electric vehicles, comprising the following steps:

[0008] Step S1: Collect the raw time series data of electric vehicle charging load, preprocess the raw time series data, and perform hybrid tokenization on the preprocessed one-dimensional time series data to obtain heterogeneous feature tokens.

[0009] Step S2: Construct a decomposition diffusion model as the generation model for electric vehicle charging load. The decomposition diffusion model performs two processes: forward noise addition and reverse noise removal. The noise addition process adds Gaussian noise to the original data step by step to generate data with different levels of noise. The reverse noise removal is performed by a noise prediction model based on the Transformer network architecture, which is used to predict and remove noise in the noisy data at each time step under the guidance of heterogeneous feature tokens, and finally restore the generated charging load data.

[0010] Step S3: Solve for the optimal parameters θ of the decomposition diffusion model using an optimization training method. * ;

[0011] Step S4: Obtain the optimal parameters θ of the model parameters θ. * Then, the charging load generation stage begins. In the charging load generation stage, the heterogeneous feature token obtained in step S1 is used as condition c and input into the decomposition diffusion model. The decomposition diffusion model outputs three types of discrete predicted values ​​for the time series: trend, seasonality, and residual.

[0012] Step S5: The three types of discrete forecast values ​​obtained in step S4—trend, seasonality, and residual—are fused and reconstructed into smooth and continuous time series, forming three continuous component series. The three continuous component series are then added together at the same time points to obtain the electric vehicle charging load curve.

[0013] In the aforementioned method for generating electric vehicle charging load, in step S1, during the hybrid tokenization process, the preprocessed one-dimensional time-series signal x is transformed from a single numerical sequence into a set of heterogeneous feature tokens containing rich structured information. The information of the three dimensions of the three types of component cue tokens is processed in parallel using three independent pipelines. The three types of component cue tokens include local multi-scale pattern tokens, global statistical feature tokens, and component-guided cue tokens: the three types of component cue tokens together constitute a set of heterogeneous feature tokens.

[0014] In the aforementioned method for generating charging load for an electric vehicle, step S1,

[0015] In the local multi-scale pattern token processing process, K different lengths {P1,…,P} are used. k ,…,P K The original time series x is segmented in parallel using sequence fragments, with each sequence fragment being segmented by a scale P corresponding to that of the sequence fragment. k The corresponding, learnable linear mapping E k Embed the sequence fragment into a dimension d model The high-dimensional feature space forms a set of sequence fragment tokens that can characterize local dynamic features at different time scales;

[0016] In the global statistical feature token processing, M global statistical features of the original time series are calculated, including mean, standard deviation, kurtosis, and skewness, forming a statistical vector s∈R. M The statistical vector s is nonlinearly transformed using a multilayer perceptron (MLP) and encoded into an independent global statistical token T containing global information. stat ;

[0017] During the component-guided cue token processing, three types of learnable component cue tokens {T} are created, corresponding to the trend, season, and residual, respectively. trend ,T season ,T resid}

[0018] In the aforementioned method for generating electric vehicle charging load, step S3 introduces an additional distillation loss L during the initial few rounds of training the decomposition-based diffusion model. distill The distillation loss term is used to constrain the trend and seasonality of the decompositional diffusion model output. It utilizes a decomposition algorithm with the results of real data decomposition as pseudo-labels to guide the decompositional diffusion model to quickly learn the basic decomposition paradigm. The decomposition algorithm employs the MSTL algorithm, and the distillation loss L... distill The mathematical expression is:

[0019] L distill =||T MSTL -T pred ||1+||S MSTL -S pred ||1 (3)

[0020] Among them, T MSTL and S MSTL These are the trend and seasonal components decomposed by the MSTL algorithm, T pred and S predThese are the trend and seasonal forecast outputs corresponding to the model, respectively.

[0021] In the aforementioned method for generating electric vehicle charging load, step S3 involves employing a composite loss function L during the training of the decomposition-based diffusion model. total End-to-end optimization of the parameters of the decomposition diffusion model is performed, whereby the end-to-end optimization refers to optimizing the composite loss function L. total The model parameters θ of the decompositional diffusion model are directly optimized, and the composite loss function L... total It is a weighted combination of constraints on multiple dimensions, expressed as:

[0022]

[0023] In the formula, L distill This is for distillation losses and is used for warm start-up. For trend smoothing loss;

[0024] Trend smoothing loss By analyzing the predicted trend component T pred Applying the second-order difference operator To achieve this, it is represented as:

[0025]

[0026] To address the seasonal consistency loss, consistency is applied to the seasonal component S at the corresponding preset period L. When the sequence length N ≤ L, circular consistency is used instead of differential alignment. The seasonal consistency loss is expressed as:

[0027]

[0028] Where S|t| represents the value of the seasonal component S at time point t, and mod is the modulo operator;

[0029] The residual whitening loss is composed of the mean penalty and the average of the absolute values ​​of the autocorrelation coefficients over multiple lags. Its purpose is to impose a whitening constraint on the residual component R, making the mean of R close to 0 and the autocorrelation coefficients over lags 1 to K close to 0. It is expressed as:

[0030]

[0031] Where ρ R (k) is the sample autocorrelation coefficient of the residual component R at lag order k, where K is the maximum lag order; α mean and α ac These are hyperparameter weight one and hyperparameter weight two, respectively, and mean(R) is the average value of the residual component R;

[0032] λ diff , λ distill , λ T , λ S and λ R These are the hyperparameter weighting coefficients for each loss.

[0033] In the aforementioned method for generating electric vehicle charging load, in step S4, the diffusion process in the decomposition diffusion model includes two processes: forward noise addition and reverse noise removal.

[0034] The positive noise addition process gradually degrades the noiseless data x0 into different noisy data x according to a preset noise schedule. t This constructs data-noise pairs (x) for training. t ,∈), and provides a supervision signal for the subsequent inverse denoising network; wherein the cumulative product calculated according to the preset noise scheduling table is Controlling the signal-to-noise ratio at each step, the forward noise addition stage is represented as:

[0035]

[0036] Where x0 is the noise-free raw data, t is the discrete diffusion time step, ∈ ~N(0,I) is standard Gaussian noise, and I is the identity matrix. It is the cumulative product calculated based on the preset noise scheduling table, representing the ratio of control signal to noise, x. t It is noisy data at time step t;

[0037] The reverse denoising process starts from the standard Gaussian noise distribution x T Starting from ~N(0,I), guided by the conditional input c, the noise prediction network ∈ θ (·) Estimate and remove noise at each time step t to obtain noisy data x. t The time step t ranges from T to 0. When t is 0, the reconstructed noise-free original data x0 is obtained.

[0038] In the aforementioned method for generating electric vehicle charging load, in step S4, the objective of training the decomposition diffusion model is the basic diffusion loss L. diff , represented as

[0039]

[0040] In the formula, the noisy data x t Obtained from the forward noise addition process, This represents the expectation of the joint distribution of the noise-free original data x0 and Gaussian noise ∈ over a diffusion time step t. Let w(t) be the squared norm 2, and w(t) be the loss weight function dependent on the signal-to-noise ratio. θ (xt ,t,c) represents the noise prediction given by the noise prediction network for the decomposed diffusion model under condition c.

[0041] In the aforementioned method for generating electric vehicle charging load, step S5 involves performing a weighted overlap-add fusion reconstruction on each component during the fusion reconstruction process, including:

[0042] For any time point j in the sequence, the reconstructed value x' j The result is obtained by dynamically weighting and summing the predicted values ​​of all sequence segments covering that point based on the confidence levels of the model output:

[0043]

[0044] In the formula, P j It is the set of all sequence segments that cover time point j. It is the predicted value of sequence segment i at time point j, with weight w. i,j The confidence level is dynamically calculated using the Softmax function based on the output confidence level of the decomposed diffusion model, and is represented as w. i,j =Softmax i (s i,j );

[0045] Finally, physical constraints are applied to the electric vehicle charging load curve. These physical constraints include a non-negativity constraint that ensures the charging load power value is always non-negative and a power boundary constraint that ensures the load curve conforms to the physical upper limit of the charging facility. The final electric vehicle charging load curve is then output.

[0046] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to perform the steps of the method described above.

[0047] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0048] The beneficial effects achieved by this invention are as follows: The method of this invention constructs a model for generating electric vehicle charging load curves through an innovative hybrid representation input and decompositional generation mechanism, overcoming the "black box" problem of traditional generation models. The model explicitly decomposes the generated load curve into three components with clear physical meaning: trend, seasonality, and residuals, thereby achieving structural interpretability. Through a hybrid tokenization method, the model can simultaneously perceive global context and multi-scale local patterns from a single time series and be explicitly guided to perform decomposition tasks, greatly improving learning efficiency and generation quality, and achieving comprehensive information representation of charging load data. Addressing the generation uncertainty problem of the model, this invention utilizes multiple loss function constraints to ensure that the generated charging curves are highly similar to real data in statistical characteristics, time-domain morphology, and frequency-domain features, achieving high fidelity and realism. This method provides a foundation for applications such as charging load forecasting, power grid planning, and charging infrastructure layout, achieving optimal allocation of charging load. Attached Figure Description

[0049] Figure 1 This is a flowchart of the electric vehicle charging load generation method in Embodiment 1 of the present invention;

[0050] Figure 2 This is a schematic diagram of the hybrid tokenization process in Embodiment 1 of the present invention;

[0051] Figure 3 This is a schematic diagram of the decomposition-based diffusion model architecture in Embodiment 1 of the present invention;

[0052] Figure 4 This is an example diagram of the charging load and decomposition components generated in Embodiment 1 of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0054] Example 1

[0055] like Figure 1 As shown, this embodiment provides a method for generating electric vehicle charging load data, which generates charging load data with high fidelity, diversity, and interpretability, including the following steps:

[0056] Step S1: Collect raw time series data of electric vehicle charging load x∈R L Where L is the sequence length and R represents the real number field, the original time series data is preprocessed, and the preprocessed one-dimensional time series data is subjected to hybrid tokenization to obtain heterogeneous feature tokens.

[0057] The preprocessing includes:

[0058] Fill in missing values ​​using interpolation;

[0059] The Z-Score normalization method is used to normalize the data to eliminate the influence of units and enhance the stability of model training.

[0060] In the hybrid tokenization process, the preprocessed one-dimensional time-series signal x is transformed from a single numerical sequence into a set of heterogeneous feature tokens containing rich structured information. For example... Figure 2 As shown, three independent pipelines are used to process information from three dimensions of three types of component cue tokens in parallel. The three types of component cue tokens include local multi-scale pattern tokens, global statistical feature tokens, and component guidance cue tokens. These three types of component cue tokens together constitute a set of heterogeneous feature tokens, which serve as the conditional input for the subsequent decompositional diffusion model, providing a structured information foundation for the generation task.

[0061] In the local multi-scale mode token processing, in order to simultaneously capture the short-term drastic fluctuations and long-term variation patterns of the charging load, K different lengths {P1,…,P} are used. k ,…,P K The original time series x is segmented in parallel using sequence fragments, with each sequence fragment being segmented by a scale P corresponding to that of the sequence fragment. k The corresponding, learnable linear mapping E k Embed the sequence fragment into a dimension d model The high-dimensional feature space forms a set of sequence fragment tokens that can characterize local dynamic features at different time scales.

[0062] In the global statistical feature token processing, to provide the decompositional diffusion model with macroscopic contextual information such as the overall location, scale, and distribution pattern of the entire time series, M global statistical features of the original time series are calculated, including mean, standard deviation, kurtosis, and skewness, forming a statistical vector s∈R. M The statistical vector s is nonlinearly transformed using a multilayer perceptron (MLP) and encoded into an independent global statistical token T containing global information. stat .

[0063] In the component-guided cue token processing, to achieve explicit control and decomposition of the data generation process, three types of learnable component cue tokens {T} are created, corresponding to trend, season, and residual, respectively. trend ,T season ,T resid The component hint token serves as a conditional instruction for the decompositional diffusion model in subsequent steps, controlling the decompositional generation process of the model so that it only processes the target component indicated by the token, including three categories: trend, seasonality, and residual.

[0064] Step S2: Construct a decompositional diffusion model as the generation model for electric vehicle charging load. This model performs two processes: forward noise addition and reverse denoising. The noise addition process adds Gaussian noise to the original data step-by-step, generating data with varying levels of noise. The reverse denoising is performed by a noise prediction model based on a Transformer network architecture. Guided by heterogeneous feature tokens, this model predicts and removes noise from the noisy data at each time step, ultimately restoring the generated charging load data. The Transformer network architecture in the decompositional diffusion model includes a self-attention mechanism and a feedforward neural network. First, the self-attention mechanism evaluates the mutual importance of all heterogeneous feature tokens to aggregate global information. Then, the feedforward neural network performs nonlinear deep processing on the aggregated information to more accurately predict noise.

[0065] like Figure 3 As shown, in the decompositional diffusion model, the heterogeneous feature token set c = {E} generated in step S1 is... k (x patch )} k ∪{T stat}∪{T trend ,T season ,T resid} is used as a conditional input. Where, {E k (x patch )} k The pattern is a local multi-scale token, representing k sequence fragments x of different lengths. patch After their respective linear mappings E k Encoded feature tokens are used to capture local dynamic features of the charging load; {T stat} is the global statistical feature token, which is encoded by a multilayer perceptron (MLP) using global statistics such as the mean, standard deviation, kurtosis, and skewness of the original time series, providing macroscopic contextual information for the model; {T trend ,T season ,T resid} is a learnable component-guided cue token used to generate cuees for trend, seasonality, and residual, guiding and controlling the decompositional generation process of the decompositional diffusion model.

[0066] Step S3: Solve for the optimal parameters θ of the decomposition diffusion model using an optimization training method. * To overcome the cold start problem of decomposition-based diffusion models simultaneously learning both the inverse denoising process and component-guided cues, this approach comprises two parts:

[0067] An additional distillation loss L is introduced in the initial few rounds of training the decomposition diffusion model. distill This is used to constrain the trend and seasonality of the decompositional diffusion model output. The distillation loss term utilizes a decomposition algorithm with the results of real data decomposition as pseudo-labels to guide the decompositional diffusion model to quickly learn the basic decomposition paradigm. The decomposition algorithm employs the MSTL algorithm (MSTL Distillation Warm-Start Method). The distillation loss L... distill The mathematical expression is:

[0068] L distill =||T MSTL -T pred ||1+||S MSTL -S pred ||1 (3)

[0069] Among them, T MSTL and S MSTL These are the trend and seasonal components decomposed by the MSTL algorithm, T pred and S pred These are the trend and seasonal forecast outputs corresponding to the model, respectively. The distillation loss L distill weight λ distill It will linearly decay from a large initial value to zero as training progresses.

[0070] The composite loss function L is used in the training process of the decomposition diffusion model. total End-to-end optimization of the parameters of the decomposed diffusion model is performed. This end-to-end optimization refers to optimizing the parameters using the composite loss function L... total The model parameters θ of the decompositional diffusion model are directly optimized. The composite loss function L... total It is a weighted combination of constraints on multiple dimensions, expressed as:

[0071]

[0072] In the formula, L distill This is for distillation losses and is used for warm start-up. Trend smoothing loss is used to ensure that the trend terms generated by the model conform to the basic physical definition of low-frequency, slowly changing trends. By analyzing the predicted trend component T pred Applying the second-order difference operator To achieve this, it is represented as:

[0073]

[0074] To address the seasonal consistency loss, consistency is applied to the seasonal component S at its corresponding preset period L. When the sequence length N ≤ L, circular consistency is used instead of differential alignment to improve the stability of short sequence scenarios. The seasonal consistency loss is expressed as:

[0075]

[0076] Where S|t| represents the value of the seasonal component S at time point t, and mod is the modulo operator, used to achieve circular consistency, so that the sequences can be connected in a circular manner when calculating periodic differences.

[0077] The residual whitening loss is composed of the mean penalty and the average of the absolute values ​​of the autocorrelation coefficients over multiple lags. Its purpose is to impose a whitening constraint on the residual component R, making the mean of R close to 0 and the autocorrelation coefficients over lags 1 to K close to 0. It is expressed as:

[0078]

[0079] Where ρ R (k) is the sample autocorrelation coefficient of the residual component R at lag order k, where K is the maximum lag order; α mean and α ac These are hyperparameter weight one and hyperparameter weight two, respectively, and mean(R) is the average value of the residual component R.

[0080] λ diff , λ distill , λ T , λ S and λ R These are the hyperparameter weighting coefficients for each loss, used to balance different optimization objectives.

[0081] Step S4: Obtain the optimal parameters θ of the model parameters θ. * Then, the charging load generation stage begins. In the charging load generation stage, the heterogeneous feature token obtained in step S1 is input as condition c, and the decomposition diffusion model outputs three types of discrete prediction values ​​for the time series: trend, seasonality, and residual.

[0082] In the decompositional diffusion model, the diffusion process includes two processes: forward noise addition and reverse noise removal.

[0083] The positive noise addition process gradually degrades the noiseless data x0 into different noisy data x according to a preset noise schedule. t This constructs data-noise pairs (x) for training. t ,∈), and provides a supervision signal for the subsequent inverse denoising network; wherein the cumulative product calculated according to the preset noise scheduling table is Controlling the signal-to-noise ratio at each step, the forward noise addition stage is represented as:

[0084]

[0085] Where x0 is the noise-free raw data, t is the discrete diffusion time step, ∈ ~N(0,I) is standard Gaussian noise, and I is the identity matrix. It is the cumulative product calculated based on the preset noise scheduling table, representing the ratio of control signal to noise, x. t It is noisy data at time step t.

[0086] The reverse denoising process starts from the standard Gaussian noise distribution x T Starting from ~N(0,I), guided by the conditional input c, the noise prediction network ∈ θ (·) Estimate and remove noise at each time step t to obtain noisy data x. t The time step t ranges from T to 0. When t is 0, the reconstructed noise-free original data x0 is obtained. The goal of training the decompositional diffusion model is to make the noise prediction network ∈ θ (·) can accurately predict the Gaussian noise ∈ added to the noise-free original data x0 at each time step t, and the objective of training the decomposition diffusion model is based on the diffusion loss L. diff , represented as

[0087]

[0088] In the formula, the noisy data x t Obtained from the forward noise addition process, This represents the expectation of the joint distribution of the noise-free original data x0 and Gaussian noise ∈ over a diffusion time step t. Let be the squared L2 norm, and w(t) be the loss weight function dependent on the signal-to-noise ratio. This represents the different levels of importance given to predictions at different noise levels during training, in order to improve the stability of model training and the final performance. θ (x t (t, c) represents the noise predictions given by the noise prediction network under condition c in the decompositional diffusion model. The noise prediction network is based on a Transformer architecture. Under the control of component guidance cue tokens, the model outputs discrete predictions of three time series components during the inference phase: Trend, Seasonality, and Residual. The three components share the same Transformer parameter θ, and the corresponding trend, seasonality, and residual sequences are generated according to the conditional input c of different component cue tokens.

[0089] Step S5: The discrete predicted values ​​of the three categories of trend, seasonality and residual obtained in step S4 are fused and reconstructed into smooth and continuous time series to form three continuous component series. Then, the three continuous component series are added together at the same time points to obtain the electric vehicle charging load curve.

[0090] During the fusion and reconstruction process, each component is fused and reconstructed using a weighted overlap-add method. For any time point j in the sequence, the reconstructed value x' j The result is obtained by dynamically weighting and summing the predicted values ​​of all sequence segments covering that point based on the confidence levels of the model output:

[0091]

[0092] In the formula, P j It is the set of all sequence segments that cover time point j. It is the predicted value of sequence segment i at time point j, with weight w. i,j The confidence level is dynamically calculated using the Softmax function based on the output confidence level of the decomposed diffusion model, and is represented as w. i,j =Softmax i (s i,j ).

[0093] Finally, physical constraints are applied to the electric vehicle charging load curve to ensure that it conforms to real physical laws. These physical constraints include a non-negativity constraint that ensures the charging load power value is always non-negative and a power boundary constraint that ensures the load curve conforms to the physical upper limit of the charging infrastructure. The final electric vehicle charging load curve is then output.

[0094] This embodiment employs a test system based on real charging data, sourced from the publicly available ACN-Data dataset containing measured records from the Jet Propulsion Laboratory (JPL) site. As a typical workplace charging scenario with 52 charging stations, the data from this site accurately reflects the charging load characteristics during weekday commutes. Data spanning over six months, encompassing thousands of complete charging sessions, was selected and processed into a total charging load time series with 15-minute intervals for model input. This time series data exhibits high nonlinearity and non-stationarity, containing not only significant daily and weekly periodicity but also reflecting long-term trend changes due to the increasing number of charging vehicles, accompanied by substantial random fluctuations. This makes the total charging load time series data obtained from the JPL site an ideal sample for evaluating model performance.

[0095] Figure 4This paper demonstrates the charging load curve and its decomposed components generated by applying the proposed method to the JPL dataset at the Jet Propulsion Laboratory site. Unlike traditional black-box models such as Generative Adversarial Networks (GANs) that can only mimic the overall shape of the load curve, this invention not only generates a total load curve that is highly consistent with real data in terms of statistical characteristics and time-frequency domain features, but also explicitly decomposes the total load curve into three components with clear physical meanings. First, the trend component smoothly reflects the long-term, slow growth of electric vehicle ownership and charging pile utilization in the working area. Second, the seasonal component clearly depicts the weekday charging pattern with a 24-hour cycle, such as the morning peak, afternoon stability, and evening trough, revealing the significant differences between weekday and weekend patterns. Finally, the residual component successfully captures random, high-frequency fluctuations that cannot be explained by trend and seasonality, such as changes caused by the irregular charging behavior of individual users. This decomposed generation method overcomes the problem of poor interpretability in existing methods. It not only ensures the authenticity and diversity of the generated data, but also represents the inherent structure of the load, such as base load and periodic shocks, thereby supporting more refined analysis and decision-making.

[0096] Example 2

[0097] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to perform the steps of the method as described in Embodiment 1.

[0098] Example 3

[0099] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method as described in Example 1.

[0100] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0101] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0102] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0103] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating charging load for electric vehicles, characterized in that, Includes the following steps: Step S1: Collect the raw time series data of electric vehicle charging load, preprocess the raw time series data, and perform hybrid tokenization on the preprocessed one-dimensional time series data to obtain heterogeneous feature tokens. Step S2: Construct a decomposition diffusion model as the generation model for electric vehicle charging load. The decomposition diffusion model performs two processes: forward noise addition and reverse noise removal. The noise addition process adds Gaussian noise to the original data step by step to generate data with different levels of noise. The reverse noise removal is performed by a noise prediction model based on the Transformer network architecture, which is used to predict and remove noise in the noisy data at each time step under the guidance of heterogeneous feature tokens, and finally restore the generated charging load data. Step S3: Solve for the optimal parameters θ of the decomposition diffusion model using an optimization training method. * ; Step S4: Obtain the optimal parameters θ of the model parameters θ * Then, the charging load generation stage begins. In the charging load generation stage, the heterogeneous feature token obtained in step S1 is used as condition c and input into the decomposition diffusion model. The decomposition diffusion model outputs three types of discrete predicted values ​​for the time series: trend, seasonality, and residual. Step S5: The three types of discrete predicted values ​​are fused and reconstructed into smooth, continuous time series to form three continuous component series. The three continuous component series are then added together at the same time points to obtain the electric vehicle charging load curve.

2. The method for generating charging load for an electric vehicle according to claim 1, characterized in that, In step S1, during the hybrid tokenization process, the preprocessed one-dimensional time-series signal x is transformed from a single numerical sequence into a set of heterogeneous feature tokens containing rich structured information. The information of the three types of component cue tokens is processed in parallel using three independent pipelines. The three types of component cue tokens include local multi-scale pattern tokens, global statistical feature tokens, and component guidance cue tokens: the three types of component cue tokens together constitute a set of heterogeneous feature tokens.

3. The method for generating charging load for an electric vehicle according to claim 2, characterized in that, In step S1, In the local multi-scale pattern token processing process, K different lengths {P1,…,P} are used. k ,…,P K The original time series x is segmented in parallel using sequence fragments, with each sequence fragment being segmented by a scale P corresponding to that of the sequence fragment. k The corresponding, learnable linear mapping E k Embed the sequence fragment into a dimension d model The high-dimensional feature space forms a set of sequence fragment tokens that can characterize local dynamic features at different time scales; In the global statistical feature token processing, M global statistical features of the original time series are calculated, including mean, standard deviation, kurtosis, and skewness, forming a statistical vector s∈R. M The statistical vector s is nonlinearly transformed using a multilayer perceptron (MLP) and encoded into an independent global statistical token T containing global information. stat ; During the component-guided cue token processing, three types of learnable component cue tokens {T} are created, corresponding to the trend, season, and residual, respectively. trend ,T season ,T resid } 4. The method for generating charging load for an electric vehicle according to claim 1, characterized in that, In step S3, an additional distillation loss L is introduced during the initial few rounds of training the decomposition diffusion model. distill The distillation loss term is used to constrain the trend and seasonality of the decompositional diffusion model output. It utilizes a decomposition algorithm with the results of real data decomposition as pseudo-labels to guide the decompositional diffusion model to quickly learn the basic decomposition paradigm. The decomposition algorithm employs the MSTL algorithm, and the distillation loss L... distill The mathematical expression is: L distill =||T MSTL -T pred ||1+||S MSTL -S pred ||1 (3) Among them, T MSTL and S MSTL These are the trend and seasonal components decomposed by the MSTL algorithm, T pred and S pred These are the trend and seasonal forecast outputs corresponding to the model, respectively.

5. The method for generating charging load for an electric vehicle according to claim 4, characterized in that, In step S3, a composite loss function L is used during the training of the decomposition diffusion model. total End-to-end optimization of the parameters of the decomposition diffusion model is performed, whereby the end-to-end optimization refers to optimizing the composite loss function L. total The model parameters θ of the decompositional diffusion model are directly optimized, and the composite loss function L... total It is a weighted combination of constraints on multiple dimensions, expressed as: In the formula, L distill This is for distillation losses and is used for warm start-up. For trend smoothing loss; Trend smoothing loss By analyzing the predicted trend component T pred Applying the second-order difference operator To achieve this, it is represented as: To address the seasonal consistency loss, consistency is applied to the seasonal component S at the corresponding preset period L. When the sequence length N ≤ L, circular consistency is used instead of differential alignment. The seasonal consistency loss is expressed as: Where S|t| represents the value of the seasonal component S at time point t, and mod is the modulo operator; The residual whitening loss is composed of the mean penalty and the average of the absolute values ​​of the autocorrelation coefficients over multiple lags. Its purpose is to impose a whitening constraint on the residual component R, making the mean of R close to 0 and the autocorrelation coefficients over lags 1 to K close to 0. It is expressed as: Where ρ R (k) is the sample autocorrelation coefficient of the residual component R at lag order k, where K is the maximum lag order; α mean and α ac These are hyperparameter weight one and hyperparameter weight two, respectively, and mean(R) is the average value of the residual component R; λ diff , λ distill , λ T , λ S and λ R These are the hyperparameter weighting coefficients for each loss.

6. The method for generating charging load for an electric vehicle according to claim 1, characterized in that, In step S4, in the decomposition diffusion model, the diffusion process includes two processes: forward noise addition and reverse noise removal. The positive noise addition process gradually degrades the noiseless data x0 into different noisy data x according to a preset noise schedule. t This constructs data-noise pairs (x) for training. t ,∈), and provides a supervision signal for the subsequent inverse denoising network; wherein the cumulative product calculated according to the preset noise scheduling table is Controlling the signal-to-noise ratio at each step, the forward noise addition stage is represented as: Where x0 is the noise-free raw data, t is the discrete diffusion time step, ∈ ~N(0,I) is standard Gaussian noise, and I is the identity matrix. It is the cumulative product calculated based on the preset noise scheduling table, representing the ratio of control signal to noise, x. t It is noisy data at time step t; The reverse denoising process starts from the standard Gaussian noise distribution x T Starting from ~N(0,I), guided by the conditional input c, the noise prediction network ∈ θ (·) Estimate and remove noise at each time step t to obtain noisy data x. t The time step t ranges from T to 0. When t is 0, the reconstructed noise-free original data x0 is obtained.

7. The method for generating charging load for an electric vehicle according to claim 1, characterized in that, In step S4, the objective of training the decomposition diffusion model is the basic diffusion loss L. diff , represented as In the formula, the noisy data x t Obtained from the forward noise addition process, This represents the expectation of the joint distribution of the noise-free original data x0 and Gaussian noise ∈ over a diffusion time step t. Let w(t) be the squared norm 2, and w(t) be the loss weight function dependent on the signal-to-noise ratio. θ (x t ,t,c) represents the noise prediction given by the noise prediction network for the decomposed diffusion model under condition c.

8. The method for generating charging load for an electric vehicle according to claim 1, characterized in that, In step S5, during the fusion reconstruction process, each component is fused and reconstructed using a weighted overlap-add method, including: For any time point j in the sequence, the reconstructed value x' j The result is obtained by dynamically weighting and summing the predicted values ​​of all sequence segments covering that point based on the confidence levels of the model output: In the formula, P j It is the set of all sequence segments that cover time point j. It is the predicted value of sequence segment i at time point j, with weight w. i,j The confidence level is dynamically calculated using the Softmax function based on the output confidence level of the decomposed diffusion model, and is represented as w. i,j =Softmax i (s i,j ); Finally, physical constraints are applied to the electric vehicle charging load curve. These physical constraints include a non-negativity constraint that ensures the charging load power value is always non-negative and a power boundary constraint that ensures the load curve conforms to the physical upper limit of the charging facility. The final electric vehicle charging load curve is then output.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method as claimed in any one of claims 1-9.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as claimed in any one of claims 1-8.

Images