CN-122026800-A - Cloud layer change simulation and photovoltaic power prediction method and related device based on relaxation time distribution and oriented to full-blind photovoltaic site

CN122026800ACN 122026800 ACN122026800 ACN 122026800ACN-122026800-A

Abstract

A cloud layer change simulation and photovoltaic power prediction method and a related device based on relaxation time distribution for a full-blind photovoltaic site are used for solving the problem of low photovoltaic power prediction precision in the absence of meteorological data. The method comprises the steps of extracting relaxation characteristics of different time scales from a photovoltaic electrical time sequence signal through relaxation time distribution analysis, establishing a mapping relation between the relaxation characteristics and cloud layer dynamic parameters, realizing indirect simulation of cloud layer change, fusing a simulation result with historical power data, inputting a deep learning model for training and prediction, and outputting a photovoltaic power point predicted value and a probability confidence interval thereof. According to the invention, cross-domain feature extraction from electrical characteristics to meteorological characteristics is realized on a data layer, and nonlinear modeling capability of deep learning is combined, so that accuracy and robustness of meteorological full-blind or weak meteorological station photovoltaic power prediction and reliability support on power grid dispatching are remarkably improved, and an innovative solution is provided for new energy consumption and optimal operation of an electric power market.

Inventors

YAO KAIWEN
QU YINPENG
CHEN HAOYANG

Assignees

湖南大学

Dates

Publication Date: 20260512
Application Date: 20260107

Claims (10)

1. The cloud layer change simulation and photovoltaic power prediction method based on relaxation time distribution for the totally blind photovoltaic site is characterized by comprising the following steps of: S1, collecting a photovoltaic time sequence signal and extracting DRT relaxation time distribution, namely collecting an electrical time sequence signal of a photovoltaic station when the photovoltaic station is in a normal operation or to-be-predicted state, processing the electrical time sequence signal by adopting a Tikhonov regularization algorithm, and extracting a relaxation time distribution function Wherein Is the relaxation time; mapping and correlating the characteristic peak of the DRT and the characteristic parameter of cloud layer change, namely, mapping the relaxation time distribution function Analyzing, extracting key parameters of DRT characteristic peak as DRT characteristic peak parameters, establishing mapping relation between DRT characteristic peak parameters and cloud layer variation characteristic parameters, and using the mapping relation to predict real-time or time period to be predicted Conversion to cloud layer variation characteristic parameter sequences ; S3, deep learning power prediction of fusion cloud layer characteristics, namely carrying out deep learning power prediction on the cloud layer change characteristic parameter sequence extracted in the step S2 With historical photovoltaic power sequences Combining time characteristics, inputting the time characteristics into a pre-trained deep learning prediction model, and outputting a photovoltaic power predicted value And probabilistic predictions thereof.
2. The cloud cover variation simulation and photovoltaic power prediction method based on DRT relaxation time distribution as set forth in claim 1, wherein the sub-step of S1 comprises the steps of: S1.1, acquiring instantaneous photovoltaic electrical measurement signals, namely acquiring high-frequency and multi-channel electrical time sequence signals through an in-station data acquisition system when a photovoltaic power station is in a normal operation or to-be-predicted state, wherein the electrical time sequence signals comprise output power of a photovoltaic array Local irradiance Ambient temperature ; S1.2, preprocessing the time sequence signals and constructing a relaxation-like process, namely denoising, normalizing and trend item separation processing are carried out on the acquired electrical time sequence signals, and the processed photovoltaic power change time sequence signals are processed Or irradiance change timing signal A response signal of a relaxation-like process with multi-time scale characteristics is regarded as to capture the dynamic response characteristic of the photovoltaic system to the rapid change of the cloud layer; S1.3, accurately extracting relaxation time Distribution (DRT), namely analyzing the time domain response signal, and converting the time domain response signal from a time domain or a frequency domain into a relaxation time distribution function by adopting a regularization inversion algorithm The function is The cloud layer change modes corresponding to different time scales can be decoupled clearly, wherein the fast scale mode corresponds to the fast movement of the cloud layer edge, and the slow scale mode corresponds to the slow drift of the large-area cloud cluster.
3. The cloud cover variation simulation and photovoltaic power prediction method based on DRT relaxation time distribution as set forth in claim 1, wherein the sub-step of S2 comprises the steps of: S2.1, DRT characteristic peak parameter extraction and quantification, namely, extracting and quantifying the characteristic peak parameter from the relaxation time distribution function Extracting and quantifying DRT characteristic peak parameters of characteristic peaks, wherein the DRT characteristic peak parameters comprise: Peak position Reflecting the time scale of the cloud layer change mode; Peak height Reflecting the intensity or energy of cloud layer change under the corresponding time scale; Peak width Reflecting the diffusivity or duration of the cloud layer change pattern; S2.2, constructing a correlation database of historical DRT characteristics and cloud layer parameters, namely constructing a DRT characteristic peak parameter set based on historical photovoltaic operation data containing synchronous meteorological data The synchronous meteorological data comprise high-resolution cloud images and sky imaging system data, and the actual cloud change characteristic parameters comprise cloud body moving speed, cloud light transmittance, cloud coverage change rate and cloud edge definition; S2.3, constructing and training a mapping model, namely constructing a machine learning mapping model, taking the DRT characteristic peak parameter extracted in the S2.1 as input, taking the corresponding actual cloud layer change characteristic parameter in a database as an output target, and performing offline supervision training, wherein the machine learning mapping model aims at learning a nonlinear mapping relation between the DRT characteristic peak parameter and a cloud layer physical change process ; S2.4, real-time cloud layer change characteristic parameter sequence The generation of DRT characteristic peak parameter set extracted in real time in the period to be predicted S2.1 is converted into cloud layer variation characteristic parameter sequence in real time by using a trained machine learning mapping model The sequence is As a priori meteorological input to the photovoltaic power prediction model.
4. The cloud cover variation simulation and photovoltaic power prediction method based on DRT relaxation time distribution as set forth in claim 1, wherein the substep of step S3 is as follows: s3.1, constructing the characteristic fusion input of the deep learning prediction model, namely, establishing the cloud layer change characteristic parameter sequence generated in the step S3 With historical photovoltaic power sequences Integrating and normalizing time characteristics to construct comprehensive input characteristic vector of deep learning prediction model ; S3.2, reasoning a deep learning prediction model optimized by a attention mechanism, namely constructing and using a pre-trained deep learning prediction model, wherein the deep learning prediction model can be used for inputting The self-adaptive weight is given to different characteristics in the system, reasoning is carried out, and the photovoltaic power predicted value of the period to be predicted is output ; And S3.3, probability prediction and uncertainty quantization, wherein the deep learning prediction model preferably adopts a quantile regression or Bayesian learning method to output a photovoltaic power predicted value and a probability density function or confidence interval of the photovoltaic power predicted value so as to quantize the uncertainty of the photovoltaic power prediction and provide a risk assessment basis for power grid dispatching decisions.
5. Cloud layer change simulation and photovoltaic power prediction device based on relaxation time distribution for full blind photovoltaic site, characterized by comprising: The signal acquisition and DRT relaxation time distribution extraction module is used for acquiring electrical time sequence signals of the photovoltaic station when the photovoltaic station is in a normal operation or to-be-predicted state, processing the electrical time sequence signals by adopting a Tikhonov regularization algorithm, and extracting a relaxation time distribution function Wherein Is the relaxation time; A DRT characteristic peak and cloud layer change characteristic parameter mapping association module for mapping the relaxation time distribution function Analyzing, extracting key parameters of DRT characteristic peak as DRT characteristic peak parameters, establishing mapping relation between DRT characteristic peak parameters and cloud layer variation characteristic parameters, and using the mapping relation to predict real-time or time period to be predicted Conversion to cloud layer variation characteristic parameter sequences ; The fused cloud characteristic deep learning power prediction module is used for extracting the cloud characteristic change parameter sequence With historical photovoltaic power sequences Combining time characteristics, inputting the time characteristics into a pre-trained deep learning prediction model, and outputting a photovoltaic power predicted value And probabilistic predictions thereof.
6. The cloud cover variation simulation and photovoltaic power prediction device based on DRT relaxation time distribution of claim 5, wherein the signal acquisition and DRT relaxation time distribution extraction module is specifically configured to: S1.1, acquiring instantaneous photovoltaic electrical measurement signals, namely acquiring high-frequency and multi-channel electrical time sequence signals through an in-station data acquisition system when a photovoltaic power station is in a normal operation or to-be-predicted state, wherein the electrical time sequence signals comprise output power of a photovoltaic array Local irradiance Ambient temperature ; S1.2, preprocessing the time sequence signals and constructing a relaxation-like process, namely denoising, normalizing and trend item separation processing are carried out on the acquired electrical time sequence signals, and the processed photovoltaic power change time sequence signals are processed Or irradiance change timing signal A response signal of a relaxation-like process with multi-time scale characteristics is regarded as to capture the dynamic response characteristic of the photovoltaic system to the rapid change of the cloud layer; S1.3, accurately extracting relaxation time Distribution (DRT), namely analyzing the time domain response signal, and converting the time domain response signal from a time domain or a frequency domain into a relaxation time distribution function by adopting a regularization inversion algorithm The function is The cloud layer change modes corresponding to different time scales can be decoupled clearly, wherein the fast scale mode corresponds to the fast movement of the cloud layer edge, and the slow scale mode corresponds to the slow drift of the large-area cloud cluster.
7. The cloud cover variation simulation and photovoltaic power prediction device based on DRT relaxation time distribution according to claim 5, wherein the DRT characteristic peak and cloud cover variation characteristic parameter mapping association module is specifically configured to: S2.1, DRT characteristic peak parameter extraction and quantification, namely, extracting and quantifying the characteristic peak parameter from the relaxation time distribution function Extracting and quantifying DRT characteristic peak parameters of characteristic peaks, wherein the DRT characteristic peak parameters comprise: Peak position Reflecting the time scale of the cloud layer change mode; Peak height Reflecting the intensity or energy of cloud layer change under the corresponding time scale; Peak width Reflecting the diffusivity or duration of the cloud layer change pattern; S2.2, constructing a correlation database of historical DRT characteristics and cloud layer parameters, namely constructing a DRT characteristic peak parameter set based on historical photovoltaic operation data containing synchronous meteorological data The synchronous meteorological data comprise high-resolution cloud images and sky imaging system data, and the actual cloud change characteristic parameters comprise cloud body moving speed, cloud light transmittance, cloud coverage change rate and cloud edge definition; S2.3, constructing and training a mapping model, namely constructing a machine learning mapping model, taking the DRT characteristic peak parameter extracted in the S2.1 as input, taking the corresponding actual cloud layer change characteristic parameter in a database as an output target, and performing offline supervision training, wherein the machine learning mapping model aims at learning a nonlinear mapping relation between the DRT characteristic peak parameter and a cloud layer physical change process ; S2.4, real-time cloud layer change characteristic parameter sequence The generation of DRT characteristic peak parameter set extracted in real time in the period to be predicted S2.1 is converted into cloud layer variation characteristic parameter sequence in real time by using a trained machine learning mapping model The sequence is As a priori meteorological input to the photovoltaic power prediction model.
8. The cloud cover variation simulation and photovoltaic power prediction device based on DRT relaxation time distribution of claim 5, wherein the fused cloud cover characteristic deep learning power prediction module is specifically used for: s3.1, constructing the characteristic fusion input of the deep learning prediction model, namely, establishing the cloud layer change characteristic parameter sequence generated in the step S3 With historical photovoltaic power sequences Integrating and normalizing time characteristics to construct comprehensive input characteristic vector of deep learning prediction model ; S3.2, reasoning a deep learning prediction model optimized by a attention mechanism, namely constructing and using a pre-trained deep learning prediction model, wherein the deep learning prediction model can be used for inputting The self-adaptive weight is given to different characteristics in the system, reasoning is carried out, and the photovoltaic power predicted value of the period to be predicted is output ; And S3.3, probability prediction and uncertainty quantization, wherein the deep learning prediction model preferably adopts a quantile regression or Bayesian learning method to output a photovoltaic power predicted value and a probability density function or confidence interval of the photovoltaic power predicted value so as to quantize the uncertainty of the photovoltaic power prediction and provide a risk assessment basis for power grid dispatching decisions.
9. A cloud layer change simulation and photovoltaic power prediction system based on relaxation time distribution for a full-blind photovoltaic site comprises a computer readable storage medium and a processor; the computer-readable storage medium is for storing executable instructions; the processor is configured to read executable instructions stored in the computer readable storage medium and execute the relaxation time distribution-based cloud layer variation simulation and photovoltaic power prediction method for the full blind photovoltaic site according to any one of claims 1 to 4.
10. A non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the relaxation time distribution based cloud cover variation simulation and photovoltaic power prediction method for a full blind photovoltaic site of any of claims 1-4.

Description

Cloud layer change simulation and photovoltaic power prediction method and related device based on relaxation time distribution and oriented to full-blind photovoltaic site Technical Field The invention belongs to the technical field of power system planning and operation, and particularly relates to a cloud layer change simulation and photovoltaic power prediction method and a related device based on relaxation time distribution and oriented to a full-blind photovoltaic site. Background The power system is undergoing a deep revolution driven by coping with climate change and energy transformation, and is mainly characterized by large-scale access and popularization of renewable energy sources with high proportion, such as wind power generation, photovoltaic power generation (PV) and the like. Photovoltaic power generation is one of the fastest growing power sources, has the advantages of cleanliness, strong deployability and the like, but the inherent intermittence and randomness of the photovoltaic power generation bring serious challenges to operation scheduling, stable control and electric market trading of electric power systems, particularly power distribution networks and regional power grids. The volatility of photovoltaic power is mainly due to the rapid local irradiance changes caused by transient cloud cover (Cloud Transients). The time scale of the fluctuation is different from a few seconds to a few minutes, the power balance and the frequency stability of the power system are directly affected, and higher requirements are placed on the standby capacity configuration of the energy storage system and the traditional unit. Therefore, the realization of high-precision and high-timeliness photovoltaic power prediction, especially for short-term (such as minutes) and ultra-short-term (such as seconds) prediction, is a key technology for ensuring safe and economic operation of a power grid. Currently, photovoltaic power prediction techniques rely mainly on the following two broad categories of methods: The method is based on historical statistics and machine learning, and is characterized in that non-meteorological features such as historical photovoltaic power data, ambient temperature and the like are directly utilized to conduct time sequence prediction (such as ARIMA, SVR (support vector machine) or LSTM (deep learning prediction model) and GRU). Although the method gets rid of absolute dependence on external weather forecast, due to the lack of cloud dynamic information which is a core physical factor causing power fluctuation, the prediction effect of the model is often limited by a repeated mode of historical data, and when sudden or atypical cloud shielding is encountered, the robustness and accuracy are drastically reduced. In addition, the existing photovoltaic power prediction method generally faces the following common limitations when meeting the challenge of high-proportion photovoltaic penetration: The limitation of the weather information deletion to the prediction precision is that for a large number of weather totally blind sites, the prediction model lacks prior information (namely cloud layer change process) for describing the physical cause of power fluctuation, so that the model can only infer based on time sequence inertia and cannot accurately capture transient power change with high fluctuation. The traditional signal analysis has insufficient resolving power for complex fluctuation, and the traditional time sequence analysis method (such as Fourier transform and wavelet analysis) is difficult to effectively decouple the change modes of different time scales (such as slow solar motion trend, medium-speed cloud cluster drift and fast cloud edge shielding) in the photovoltaic power signal in a physical interpretable way. The problem of "black box" input of a deep-learning predictive model is that although a deep-learning predictive model (e.g., LSTM) performs well in time series prediction, if features of intrinsic physical significance (e.g., cloud dynamic parameters) are lacking in the input features, the generalization ability of the model and the predictive ability for abrupt events remain insufficient. Disclosure of Invention The invention aims to solve the technical problem of providing a cloud layer change simulation and photovoltaic power prediction method and a related device based on relaxation time distribution for a full-blind photovoltaic site, which are independent of traditional external weather forecast and can indirectly and accurately extract cloud layer dynamic change characteristics from operation data of a photovoltaic system. The framework can decouple complex time sequence signals into features with physical significance and take the features as key inputs of a deep learning prediction model, so that a high-precision and high-robustness power prediction solution for a weather full-blind photovoltaic site is constructed. In order to solve the technical pro