A distributed photovoltaic power prediction short-term diffusion algorithm

Abstract

The application provides a distributed photovoltaic power prediction short-term diffusion algorithm, data preprocessing and capacity normalization, multi-modal condition coding, physical gate-based Transformer baseline prediction, condition diffusion uncertainty modeling, and physical power reconstruction. For newly added stations without historical data, only the installed capacity and local NWP input are required. The pre-trained global condition diffusion network can be reused to directly output high-precision probability prediction, perfectly meeting the business needs of distributed photovoltaic rapid expansion.

Description

Technical Field

[0001] This invention relates to the fields of new energy power generation prediction and artificial intelligence technology, and in particular to a short-term diffusion algorithm for distributed photovoltaic power prediction. Background Technology

[0002] With the rapid development of distributed photovoltaic (PV) power generation, accurate PV power prediction is crucial for the safe and stable operation of the power grid and electricity market transactions. Traditional centralized PV prediction models are often trained on a single PV site and are difficult to directly transfer to new sites. Distributed PV is characterized by a large number of PV sites, significant differences in capacity among individual sites, and the lack of historical data for some sites (cold start problem).

[0003] Existing deep learning prediction methods have the following shortcomings: Weak generalization ability: The model often "memorizes" the output curve of a single station and has difficulty sharing spatiotemporal patterns among multiple stations.

[0004] Lack of physical constraints: Pure black-box models are prone to violating common sense in physics, such as predicting tiny power at night when there is no radiation, leading to accumulated errors.

[0005] Limitations of point prediction: Traditional models mostly output deterministic point prediction results, which cannot quantify the great uncertainty brought about by meteorological conditions (such as cloud movement), and are difficult to meet the power grid's requirements for probability confidence intervals. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art by providing a distributed photovoltaic power prediction short-term diffusion algorithm that enables distributed photovoltaic power prediction with one-time training, multi-site sharing, new site extension, and probabilistic prediction.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a short-term diffusion algorithm for distributed photovoltaic power prediction, comprising the following steps: S1. Data preprocessing and capacity normalization: The data discrepancies in distributed data centers are reflected in the installed capacity, and a normalization strategy based on installed capacity is adopted: ; in, This is the measured absolute power; This refers to the upper limit of the installed capacity or power of distributed power stations; S2, Multimodal Conditional Coding; S3. Physically Gated Transformer Baseline Prediction; S4. Conditional diffusion uncertainty modeling; S5, Physical Power Reconfiguration.

[0008] Furthermore, in S2, the conditional input includes three dimensions, specifically: Site embedding: Assign a learnable vector to each discrete site ID to capture the static geographic and equipment characteristics of discrete sites; Meteorological coding: Temporal features of multi-source NWP sequences within the forecast window are extracted through multi-layer Transformer embedding layers, supplemented by residual connections and gating fusion mechanisms; Historical power coding: The historical true power sequence before the start of reporting is concatenated with the effective mask, and the historical trend bias is extracted by a multilayer perceptron to further correct the prediction.

[0009] Furthermore, S3 specifically refers to: Radiation Direct Mapping Branch: Utilizing the maximum value of shortwave radiation, the fundamental monotonic forms of radiation and power are learned through shallow MLP and sigmoid activation. ; in, The model mapping value for radiation; It is the sigmoid activation function; This represents the maximum irradiance value in multi-source meteorological data. Spatiotemporal feature extraction: The meteorological sequence, time code, station embedding and nighttime indicator variables are concatenated and fed into a multi-head self-attention Transformer to output residual correction. Nighttime physical gating: Dynamically generating nighttime masks based on radiation thresholds The final baseline output is physically reset to zero: ; in, Predict power values ​​for the model; This is a mask for nighttime.

[0010] Furthermore, S4 specifically involves introducing a diffusion model based on 1D-UNet to model the residuals: Residual definition: The target residual is defined as the difference between the true normalized power and the baseline prediction. ; in, To predict residual values; This represents the actual power output of the photovoltaic system. Forward noise addition process: During the training phase, Gaussian noise is added to the parameters step by step. : ; in, For the first Noisy samples of the step; For the first Noisy samples of the step; Inverse denoising process: U-Net uses noisy sequences Time step The conditional encoding generated by S2 is used as input to predict the injected noise. The loss function is the standard MSE: ; in, Train the loss function for the diffusion model; For the first Noisy photovoltaic residual time series samples of diffusion step; For conditional control information; Inference Sampling: During the inference phase, starting from pure Gaussian noise, U-Net is used to progressively denoise and obtain the prediction residual. The baseline is then overlaid to obtain the final normalized prediction: ; By sampling multiple times, the probability distribution and confidence interval of the naturally formed power prediction are obtained.

[0011] Furthermore, S5 specifically includes: For the normalized sequence of the predicted output Upon final delivery or application, only the installed capacity (Cap) of the target site needs to be obtained, which is then converted to the actual physical power. .

[0012] The beneficial effects of this invention are: alignment and normalization of multi-source meteorological and power data: normalize the actual power of different stations to the [0,1] interval according to their installed capacity, decoupling the absolute value and time series form of power; Historical observation feature encoding: Extract the historical measured normalized power and effective mask before the reporting, and generate a conditional bias vector; Physically constrained Transformer baseline prediction: Input multi-source meteorological sequences and station embeddings, combine nighttime gating mechanism and shortwave radiation mapping, and output deterministic normalized power baseline; Residual uncertainty modeling based on conditional diffusion model: Based on the baseline network output, a one-dimensional time series U-Net is used to diffuse and denoise the residual between the baseline and the true power under the guidance of conditional features. Inverse normalization and physical power reconstruction: Multiply the predicted normalized power by the installed capacity of the corresponding power station to restore the physical power with a probability confidence interval. Attached Figure Description

[0013] Figure 1 This is a flowchart of a short-term diffusion algorithm for distributed photovoltaic power prediction. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0015] Please see Figure 1 A short-term diffusion algorithm for distributed photovoltaic power prediction includes the following steps: S1. Data preprocessing and capacity normalization: The data discrepancies in distributed data centers are reflected in the installed capacity, and a normalization strategy based on installed capacity is adopted: ; in, This is the measured absolute power; This refers to the upper limit of the installed capacity or power of distributed power stations; S2, Multimodal Conditional Coding; S3. Physically Gated Transformer Baseline Prediction; S4. Conditional diffusion uncertainty modeling; S5, Physical Power Reconfiguration.

[0016] In S2, the conditional input includes three dimensions, specifically: Site embedding: Assign a learnable vector to each discrete site ID to capture the static geographic and equipment characteristics of discrete sites; Meteorological coding: Temporal features of multi-source NWP sequences within the forecast window are extracted through multiple Transformer coding layers, supplemented by residual connections and gating fusion mechanisms; Historical power coding: The historical true power sequence before the start of reporting is concatenated with the effective mask, and the historical trend bias is extracted by a multilayer perceptron to further correct the prediction.

[0017] Specifically, S3 is: Radiation Direct Mapping Branch: Utilizing the maximum value of shortwave radiation, the fundamental monotonic forms of radiation and power are learned through shallow MLP and sigmoid activation. ; in, The model mapping value for radiation; It is the sigmoid activation function; This represents the maximum irradiance value in multi-source meteorological data. Spatiotemporal feature extraction: The meteorological sequence, time code, station embedding and nighttime indicator variables are concatenated and fed into a multi-head self-attention Transformer to output residual correction. Nighttime physical gating: Dynamically generating nighttime masks based on radiation thresholds The final baseline output is physically reset to zero: ; in, Predict power values ​​for the model; This is a mask for nighttime.

[0018] Specifically, S4 involves introducing a diffusion model based on 1D-UNet to model the residuals. Residual definition: The target residual is defined as the difference between the true normalized power and the baseline prediction. ; in, To predict residual values; This represents the actual power output of the photovoltaic system. Forward noise addition process: During the training phase, Gaussian noise is added to the parameters step by step. : ; in, For the first Noisy samples of the step; For the first Noisy samples of the step; Inverse denoising process: U-Net uses noisy sequences Time step The conditional encoding generated by S2 is used as input to predict the injected noise. The loss function is the standard MSE: ; in, Train the loss function (MSE mean squared loss) for the diffusion model; No. Noisy photovoltaic residual time series samples of diffusion step; For conditional control information; Inference Sampling: During the inference phase, starting from pure Gaussian noise, U-Net is used to progressively denoise and obtain the prediction residual. The baseline is then overlaid to obtain the final normalized prediction: ; By sampling multiple times, the probability distribution and confidence interval of the naturally formed power prediction are obtained.

[0019] Specifically, S5 is: For the normalized sequence of the predicted output Upon final delivery or application, only the installed capacity (Cap) of the target site needs to be obtained, which is then converted to the actual physical power. .

[0020] Driven by both physics and data: A pioneering approach that rigidly injects a "shortwave radiation mapping branch" and a "nighttime physical gating" mechanism into the deep Transformer backbone. This effectively eliminates non-physical violations in boundary conditions (such as nighttime) of purely data-driven models.

[0021] The probabilistic prediction architecture for residual diffusion: Unlike directly using a diffusion model to generate absolute power values, this algorithm utilizes PvMeteoTimeTransformer to generate a high-precision normalized baseline, and uses a 1D U-Net diffusion model to learn only the "prediction error (residual)" between the baseline and the true value. This design significantly reduces the generation difficulty and endows the model with excellent probability interval estimation capabilities while ensuring deterministic accuracy. Zero-shot / few-shot transfer capability: Based on a strict installed capacity normalization mechanism and a decoupled design for site embedding. For new sites without historical data, only their installed capacity and local NWP input are required. By reusing the pre-trained global conditional diffusion network, high-precision probability predictions can be directly output, perfectly meeting the business needs of rapid expansion of distributed photovoltaic power.

[0022] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be defined by the appended claims.

Claims

1. A short-term diffusion algorithm for distributed photovoltaic power prediction, characterized in that, Includes the following steps: S1. Data preprocessing and capacity normalization: The data discrepancies in distributed data centers are reflected in the installed capacity, and a normalization strategy based on installed capacity is adopted: ； in, This is the measured absolute power; This refers to the upper limit of the installed capacity or power of distributed power stations; S2, Multimodal Conditional Coding; S3. Physically Gated Transformer Baseline Prediction; S4. Conditional diffusion uncertainty modeling; S5, Physical Power Reconfiguration.

2. The distributed photovoltaic power prediction short-term diffusion algorithm according to claim 1, characterized in that, In S2, the conditional input includes three dimensions, specifically: Site embedding: Assign a learnable vector to each discrete site ID to capture the static geographic and equipment characteristics of discrete sites; Meteorological coding: Temporal features of multi-source NWP sequences within the forecast window are extracted through multiple Transformer coding layers, supplemented by residual connections and gating fusion mechanisms; Historical power coding: The historical true power sequence before the start of reporting is concatenated with the effective mask, and the historical trend bias is extracted by a multilayer perceptron to further correct the prediction.

3. The distributed photovoltaic power prediction short-term diffusion algorithm according to claim 2, characterized in that, Specifically, S3 is: Radiation Direct Mapping Branch: Utilizing the maximum value of shortwave radiation, the fundamental monotonic forms of radiation and power are learned through shallow MLP and sigmoid activation. ； in, The model mapping value for radiation; It is the sigmoid activation function; This represents the maximum irradiance value in multi-source meteorological data. Spatiotemporal feature extraction: The meteorological sequence, time code, station embedding and nighttime indicator variables are concatenated and fed into a multi-head self-attention Transformer to output residual correction. Nighttime physical gating: Dynamically generating nighttime masks based on radiation thresholds The final baseline output is physically reset to zero: ； in, Predict power values ​​for the model; This is a mask for nighttime.

4. The distributed photovoltaic power prediction short-term diffusion algorithm according to claim 3, characterized in that, Specifically, S4 involves introducing a diffusion model based on 1D-UNet to model the residuals. Residual definition: The target residual is defined as the difference between the true normalized power and the baseline prediction. ； in, To predict residual values; This represents the actual power output of the photovoltaic system. Forward noise addition process: During the training phase, Gaussian noise is added to the parameters step by step. : ； in, For the first Noisy samples of the step; For the first Noisy samples of the step; Inverse denoising process: U-Net uses noisy sequences Time step The conditional encoding generated by S2 is used as input to predict the injected noise. The loss function is the standard MSE: ； in, Train the loss function for the diffusion model; For the first Noisy photovoltaic residual time series samples of diffusion step; For conditional control information; Inference Sampling: During the inference phase, starting from pure Gaussian noise, U-Net is used to progressively denoise and obtain the prediction residual. The baseline is then overlaid to obtain the final normalized prediction: ； By sampling multiple times, the probability distribution and confidence interval of the naturally formed power prediction are obtained.

5. The distributed photovoltaic power prediction short-term diffusion algorithm according to claim 4, characterized in that, Specifically, S5 is: For the normalized sequence of the predicted output Upon final delivery or application, only the installed capacity (Cap) of the target site needs to be obtained, which is then converted to the actual physical power. 。

Images