A load curve generation method and system based on coarse-to-fine flow matching

Abstract

The application discloses a load curve generation method and system based on coarse-to-fine flow matching, and the method comprises the following steps: step S1, acquiring load power time series data and performing pretreatment; step S2, for each pretreated daily load curve, performing multi-resolution time hierarchy construction; step S3, training an unconditional velocity field model based on the probability path between prior noise and the real coarse resolution curve at the coarsest resolution layer; step S4, training a conditional velocity field model with the corresponding coarse resolution curve as the condition at the refining layer; step S5, constructing a terminal load curve estimation based on the velocity field output, and constraining the model by using a multi-granularity field perception loss; and step S6, generating a load curve by using the trained unconditional velocity field model. The application reduces the learning difficulty of single-stage high-dimensional time series generation, and takes into account the global trend and local fluctuation.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power system load modeling and artificial intelligence generation model, specifically involving a load curve generation method and system based on coarse-to-fine flow matching. Background Technology

[0002] AI data centers undertake tasks such as artificial intelligence training, inference, or high-performance computing. Their load curves are jointly affected by GPU / AI accelerator utilization, task scheduling, user request intensity, and cooling system dynamics. How to generate high-resolution load curves is a key problem that needs to be solved in power grid planning and load forecasting.

[0003] Currently, given the severe scarcity of real-world load data for AI data centers due to privacy restrictions and data acquisition costs, there is a lack of design for multi-time-resolution structures of load curves. This makes learning difficult when generating high-dimensional time series in a single stage, and it is also difficult to take into account both global trends and local fluctuations. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a load curve generation method and system based on coarse-to-fine flow matching. The unconditional flow matching model is trained at the lowest time resolution to learn the global load trend. Then, at each higher resolution layer, the conditional flow matching model is trained with the previous coarse resolution trajectory as a condition, and local fluctuations and sudden changes are recovered step by step. This reduces the learning difficulty of single-stage high-dimensional time series generation and takes into account both global trends and local fluctuations.

[0005] The objective of this invention is achieved through the following technical solution: a method for generating load curves based on coarse-to-fine flow matching, comprising the following steps:

[0006] Step S1: Obtain load power time series data and perform preprocessing;

[0007] Step S2: For each preprocessed daily load curve, construct a multi-resolution time hierarchy;

[0008] Step S3: Train the unconditional velocity field model at the coarsest resolution layer based on the probability path between prior noise and the true coarse resolution curve;

[0009] Step S4: Train the conditional velocity field model in the refinement layer using the corresponding coarse resolution curve as a condition;

[0010] Step S5: Construct the terminal load curve estimate based on the velocity field output, and use multi-granularity domain perception loss to constrain the model;

[0011] Step S6: Generate the load curve using the trained unconditional velocity field model.

[0012] A load curve generation system based on coarse-to-fine flow matching includes:

[0013] The data acquisition and preprocessing module acquires load power time series data and performs preprocessing.

[0014] The multi-resolution construction module constructs a multi-resolution time hierarchy for each preprocessed daily load curve;

[0015] The coarsest-grained unconditional flow matching training module trains the unconditional velocity field model at the coarsest layer based on the probability path between prior noise and the real coarse-resolution curve.

[0016] The refined layer conditional flow matching training module trains the conditional velocity field model in the refined layer using the corresponding coarse resolution curve as a condition.

[0017] The terminal estimation and multi-granularity domain perception loss calculation module constructs the terminal load curve estimate based on the velocity field output and uses multi-granularity domain perception loss to constrain the model;

[0018] The load generation module uses a trained unconditional velocity field model to generate load curves.

[0019] The beneficial effects of the present invention are: (1) the high-resolution load curve generation task is decomposed into a continuous probability transmission task under multiple time resolutions; firstly, an unconditional flow matching model is trained at the lowest time resolution to learn the global load trend; then, at each higher resolution layer, a conditional flow matching model is trained with the previous coarse resolution trajectory as a condition to recover local fluctuations and sudden changes step by step; this mechanism reduces the learning difficulty of single-stage high-dimensional time series generation and takes into account both global trends and local fluctuations.

[0020] (2) During the flow matching training process, the terminal load curve is estimated based on the velocity field output (instead of full ODE sampling), and the cross-scale aggregation consistency loss, maximum ramp rate loss and first-order difference consistency loss are calculated on the estimated curve. This allows for explicit constraints on the cross-scale consistency and physical feasibility of the generated curve during the training phase, avoiding the high computational cost of full sampling every time during training.

[0021] (3) In the refinement layer, a conditional encoder is introduced to encode the coarse-resolution load trajectory into conditional features, which are then input into the velocity field network along with the current state and time embedding, thereby achieving fine-resolution conditional generation based on coarse-grained trends. The velocity network in the refinement layer does not need to relearn the global trend from noise; it only needs to supplement local high-frequency fluctuations under coarse-grained constraints, thus reducing the training difficulty. Attached Figure Description

[0022] Figure 1This is a flowchart of the method of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0024] like Figure 1 As shown, a method for generating load curves based on coarse-to-fine flow matching is characterized by the following steps:

[0025] Step S1: Obtain load power time series data and perform preprocessing;

[0026] Step S1 includes:

[0027] S101. Obtain load power time series data according to a preset time interval as raw data;

[0028] S102. Perform preprocessing on the raw data, including time alignment, missing value imputation, outlier detection and handling, and normalization; divide the continuous sequence into daily load curve samples of fixed length according to the preset time window;

[0029] S103. Obtain multiple daily load curve samples according to steps S101~S102 to form a sample set. ;in This represents the sample of the i-th daily load curve. =1,2,…,N, where the time length of each daily load curve is T, and N represents the number of daily load curve samples in the sample set. The time length refers to the number of sampling points for each daily load curve.

[0030] Step S2: For each preprocessed daily load curve, construct a multi-resolution time hierarchy;

[0031] Step S2 includes:

[0032] For each daily load curve sample in the sample set Multi-resolution time hierarchy is constructed through iterative downsampling: ;

[0033] Among them, the length of each layer is full ,in The highest resolution, i.e., the original resolution. For the lowest sampling resolution, downsampling is preferably average pooling, with a certain number of layers. It can be flexibly configured according to the characteristics of the data.

[0034] Step S3: Train the unconditional velocity field model at the coarsest resolution layer based on the probability path between prior noise and the true coarse resolution curve;

[0035] S301. For the coarsest resolution layer, define the linear interpolation probability path at time t. :

[0036] ;

[0037] in, , Sampling prior distribution The prior distribution is a standard Gaussian distribution. , For the true coarsest resolution load curve ;

[0038] S302. For linear interpolation probability paths Terminal estimation of the coarsest resolution layer is constructed using the velocity field output. :

[0039]

[0040] S303. Training Neural Velocity Field Model This model is implemented using a neural network, and the targeting speed is... The velocity regression loss function is used:

[0041] ;

[0042] By minimizing To match the real velocity field; and simultaneously apply the maximum gradeability loss. And first-order difference consistency loss :

[0043] ;

[0044] ;

[0045] in, The maximum ramp rate threshold for the coarsest resolution layer; The maximum ramp rate for generating curves at the coarsest resolution layer:

[0046] Terminal estimation for the coarsest resolution layer Calculate adjacent sampling points , The first difference between them is divided into :

[0047] ;

[0048] This indicates the time interval for sampling at the coarsest resolution layer;

[0049] Maximum gradeability:

[0050]

[0051] , These are the first-order differences between the terminal estimation and the true curve at the coarsest resolution layer, respectively.

[0052] The total loss of the coarsest layer is:

[0053] .

[0054] Step S4: Train the conditional velocity field model in the refinement layer using the corresponding coarse resolution curve as a condition;

[0055] Step S4 includes:

[0056] S401. For each refining tier The actual load trajectory with a coarser resolution via condition encoder Mapping to compact conditional features The refined layer refers to all resolution layers other than the coarsest resolution layer;

[0057] S402. Construct the refining layer using a linear interpolation probability path. Interpolation state:

[0058] , ;

[0059] in, For prior noise, For the refining layer The true resolution load curve;

[0060] Conditional velocity field model The input is the current interpolation state. Sinusoidal time embedding TE(t) and conditional features The splicing is performed, and the training objective is still velocity regression loss. During training, the real coarse resolution curve is replaced with the curve generated by the previous stage model or the curve with added perturbation as the conditional input to improve the robustness of the conditional velocity field model to the generation error in the inference stage.

[0061] Step S5: Construct the terminal load curve estimate based on the velocity field output, and use multi-granularity domain perception loss to constrain the model;

[0062] Step S5 includes:

[0063] S501. During the training phase, the interpolated state... Terminal estimation is constructed using the velocity field output. :

[0064]

[0065] in, The terminal estimate of the complete load curve is obtained directly from the single-step output of the velocity field, without the need for a complete ODE solution;

[0066] S502. Based on terminal estimation, for the refining layer Calculate the cross-scale aggregation consistency loss:

[0067] ;

[0068] in This is a downsampling operator that ensures the generated fine-resolution curve can be downsampled to recover the corresponding coarse-resolution curve.

[0069] S503. Calculate the maximum gradeability loss:

[0070] Definition of the first In layer terminal estimation, the first-order difference between adjacent sampling points is divided into:

[0071] ;

[0072] Indicates the first The sampling time interval of a layer (depending on the layer resolution)

[0073] Maximum gradeability:

[0074] ;

[0075] The maximum gradeability loss is:

[0076] ;

[0077] in For the first The maximum ramp rate threshold for a layer can be set based on the maximum ramp rate of real training samples, a preset quantile (such as 95% or 99%), or the physical ramp rate capability of the data center's power supply and cooling system.

[0078] S504: The curve generated using L1 distance constraints has similar first-order difference characteristics to the true curve (L2 distance, Wasserstein distance, or MMD distance can also be used):

[0079] ;

[0080] in This represents a first-order difference operator.

[0081] S505. Determine the total loss of the refining layer as follows:

[0082] .

[0083] in , , λ These are the weighting coefficients for each loss term;

[0084] Step S6: Generate the load curve using the trained unconditional velocity field model.

[0085] Step S6 includes:

[0086] S601. Sample the initial state of the coarsest resolution layer from the prior distribution. ;

[0087] S602. Utilizing a trained unconditional velocity field model Integrating along the continuous time path from t=0 to t=1 using a numerical integrator (such as Euler, RK4, or Dopri5 ordinary differential equation solvers):

[0088] ;

[0089] Integrating to obtain the coarsest resolution generated curve ;

[0090] S603. For each refining layer : :by For conditions (input it into the condition encoder) Obtain conditional features Using the conditional velocity field model Solving using ODE:

[0091]

[0092] Points are used to obtain the refinement layer generation curve ;

[0093] S604. Final output: Complete daily load curve at original resolution. Then, inverse normalization is performed to obtain the load curve under the actual power unit.

[0094] In embodiments of this application, the generated synthetic load curve dataset can also be used to train a downstream load prediction model. TRTR (training on real data, testing on real data) and TSTR (training on synthetic data, testing on real data) can be compared to evaluate the actual predictive utility of the synthetic data.

[0095] A load curve generation system based on coarse-to-fine flow matching includes:

[0096] The data acquisition and preprocessing module acquires load power time series data and performs preprocessing.

[0097] The multi-resolution construction module constructs a multi-resolution time hierarchy for each preprocessed daily load curve;

[0098] The coarsest-grained unconditional flow matching training module trains the unconditional velocity field model at the coarsest layer based on the probability path between prior noise and the real coarse-resolution curve.

[0099] The refined layer conditional flow matching training module trains the conditional velocity field model in the refined layer using the corresponding coarse resolution curve as a condition.

[0100] The terminal estimation and multi-granularity domain perception loss calculation module constructs the terminal load curve estimate based on the velocity field output and uses multi-granularity domain perception loss to constrain the model.

[0101] The load generation module uses a trained unconditional velocity field model to generate load curves.

[0102] In summary, this invention decomposes high-dimensional generation into progressively lower-dimensional tasks, reducing learning difficulty and balancing global trends with local fluctuations. Physical constraints can be applied during training without requiring complete ODE sampling, balancing training efficiency and generation quality. Multi-granularity constraints are employed: cross-aggregation ensures cross-scale consistency; maximum ramp rate prevents unrealistic jumps; first-order difference constraints guarantee the authenticity of change patterns between adjacent time steps; and the refinement layer, conditioned on coarse-grained trends, only needs to learn local high-frequency details, improving the authenticity of fine-grained curves.

[0103] Experimental results on multiple real-world datasets demonstrate that our proposed method outperforms or matches the performance of VAE / GAN / WGAN / diffusion models and standard flow matching in most statistical fidelity metrics and downstream prediction tasks, particularly excelling in KL divergence, marginal distribution difference, peak-to-average ratio difference, and downstream prediction RMSE. Compared to diffusion models and GANs, it offers an advantage in training time (approximately 79 seconds vs. diffusion 223 seconds / GAN 346 seconds). Although multi-stage refinement leads to a longer generation time than single-stage models, generating 10K samples still only requires seconds (approximately 4 seconds), demonstrating a good quality-efficiency trade-off.

[0104] The foregoing description illustrates and describes a preferred embodiment of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for generating load curves based on coarse-to-fine flow matching, characterized in that: Includes the following steps: Step S1: Obtain load power time series data and perform preprocessing; Step S2: For each preprocessed daily load curve, construct a multi-resolution time hierarchy; Step S3: Train the unconditional velocity field model at the coarsest resolution layer based on the probability path between prior noise and the true coarse resolution curve; Step S4: Train the conditional velocity field model in the refinement layer using the corresponding coarse resolution curve as a condition; Step S5: Construct the terminal load curve estimate based on the velocity field output, and use multi-granularity domain perception loss to constrain the model; Step S6: Generate the load curve using the trained unconditional velocity field model.

2. The method for generating a load curve based on coarse-to-fine flow matching according to claim 1, characterized in that: Step S1 includes: S101. Obtain load power time series data according to a preset time interval as raw data; S102. Perform preprocessing on the raw data, including time alignment, missing value imputation, outlier detection and handling, and normalization; then, divide the preprocessed continuous sequence into daily load curve samples of fixed length according to the preset time window. S103. Obtain multiple daily load curve samples according to steps S101~S102 to form a sample set. ;in This represents the sample of the i-th daily load curve. =1,2,…,N, where the time length of each daily load curve is T, and N represents the number of daily load curve samples in the sample set. The time length refers to the number of sampling points for each daily load curve.

3. The method for generating a load curve based on coarse-to-fine flow matching according to claim 1, characterized in that: Step S2 includes: For each daily load curve sample in the sample set Multi-resolution time hierarchy is constructed through iterative downsampling: ； Among them, the length of each layer satisfies ,in The highest resolution, i.e., the original resolution. For the lowest sampling resolution, Indicates the resolution of the l-th layer. Let l represent the length of the l-th layer, where l = 1, 2, ..., L.

4. The method for generating a load curve based on coarse-to-fine flow matching according to claim 1, characterized in that: Step S3 includes: S301. For the coarsest resolution layer, define the linear interpolation probability path at time t. : ； in, , Sampling prior distribution The prior distribution is a standard Gaussian distribution. , For the true coarsest resolution load curve ; S302. For linear interpolation probability paths Terminal estimation of the coarsest resolution layer is constructed using the velocity field output. : ； S303. Training Neural Velocity Field Model This model is implemented using a neural network, and the targeting speed is... The velocity regression loss function is used: ； By minimizing To match the real velocity field; and simultaneously apply the maximum gradeability loss. And first-order difference consistency loss : ； ； in, The maximum ramp rate threshold for the coarsest resolution layer; The maximum ramp rate for generating curves at the coarsest resolution layer: Terminal estimation for the coarsest resolution layer Calculate adjacent sampling points , The first difference between them is divided into : ； This indicates the time interval for sampling at the coarsest resolution layer; Maximum gradeability: ； , These are the first-order differences between the terminal estimation and the true curve at the coarsest resolution layer, respectively. The total loss of the coarsest layer is: 。 5. The method for generating a load curve based on coarse-to-fine flow matching according to claim 1, characterized in that: Step S4 includes: S401. For each refining tier The true load trajectory with a coarser resolution via condition encoder Mapping to compact conditional features The refined layer refers to all resolution layers other than the coarsest resolution layer; S402. Construct the refining layer using a linear interpolation probability path. Interpolation state: ， ； in, For prior noise, For the refining layer The true resolution load curve; Conditional velocity field model The input is the current interpolation state. Sinusoidal time embedding TE(t) and conditional features The splicing process still aims to train for velocity regression loss. The training data consists of all daily load curve samples in the sample set at a resolution of [resolution value missing]. For the data below, during training, the real coarse resolution curve is replaced with the curve generated by the previous stage model or the curve with added perturbation as the conditional input to improve the robustness of the conditional velocity field model to the generation error in the inference stage.

6. The method for generating a load curve based on coarse-to-fine flow matching according to claim 1, characterized in that: Step S5 includes: S501. During the training phase, the interpolated state... Terminal estimation is constructed using the velocity field output. : ； in, This is the terminal estimate of the complete load curve obtained directly from the single-step output of the velocity field; S502. Based on terminal estimation, for the refining layer Calculate the cross-scale aggregation consistency loss: ； in This is a downsampling operator that ensures the generated fine-resolution curve can be downsampled to recover the corresponding coarse-resolution curve. S503. Calculate the maximum gradeability loss: Definition of the first In layer terminal estimation, the first-order difference between adjacent sampling points is divided into: ； Indicates the first The sampling time interval of the layer; Maximum gradeability: ； The maximum gradeability loss is: ； in For the first Maximum ramp rate threshold for the layer; S504: The curve generated using L1 distance constraints has similar first-order difference characteristics to the true curve. ； in This represents a first-order difference operator; S505. Determine the total loss of the refining layer as follows: ； in , , λ These are the weighting coefficients for each loss term.

7. The method for generating a load curve based on coarse-to-fine flow matching according to claim 1, characterized in that: Step S6 includes: S601. Sample the initial state of the coarsest resolution layer from the prior distribution. ; S602. Utilizing a trained unconditional velocity field model Integrating along the continuous time path from t=0 to t=1 using a numerical integrator: ； Integrating to obtain the coarsest resolution generated curve ; S603. For each refining layer : :by As a condition, input it into the condition encoder. Obtain conditional features Using the conditional velocity field model Solving using ODE: ； Points are used to obtain the refinement layer generation curve ; S604. Final output: Complete daily load curve at original resolution. Then, inverse normalization is performed to obtain the load curve under the actual power unit.

8. A load curve generation system based on coarse-to-fine flow matching, wherein the method described in any one of claims 1 to 7 is characterized in that: include: The data acquisition and preprocessing module acquires load power time series data and performs preprocessing. The multi-resolution construction module constructs a multi-resolution time hierarchy for each preprocessed daily load curve; The coarsest-grained unconditional flow matching training module trains the unconditional velocity field model at the coarsest layer based on the probability path between prior noise and the real coarse-resolution curve. The refined layer conditional flow matching training module trains the conditional velocity field model in the refined layer using the corresponding coarse resolution curve as a condition. The terminal estimation and multi-granularity domain perception loss calculation module constructs the terminal load curve estimate based on the velocity field output and uses multi-granularity domain perception loss to constrain the model. The load generation module uses a trained unconditional velocity field model to generate load curves.

Images