CN-122026309-A - Probabilistic photovoltaic power prediction method and system based on parameter sensitivity

CN122026309ACN 122026309 ACN122026309 ACN 122026309ACN-122026309-A

Abstract

The invention discloses a probabilistic photovoltaic power prediction method and a probabilistic photovoltaic power prediction system based on parameter sensitivity, and relates to the field of renewable energy power prediction. Firstly, an integrated physical chain comprising incidence angle/spectrum correction and thermal model is constructed, weather models are divided according to clear sky indexes, a computable lower bound of power prediction variance is obtained through first-order sensitivity propagation, so that baseline deterministic prediction with deviation correction is obtained, then, a state-aware gradient lifting tree model is introduced to learn and calibrate NWP errors and conditional variances under different weather models, the conditional covariance is transmitted to a power side, and calibrated probability distribution and prediction intervals are generated. During operation, the high-precision deterministic prediction through deviation correction is combined with the variance estimation of the state condition to form a reliable probabilistic result of coverage rate and width matching, the accuracy of deterministic prediction is improved, an interpretable and calibratable probability interval is provided, and the capability of tracking abrupt changes such as cloud-induced climbing/slumping is enhanced.

Inventors

LI RUN
NA GUANGYU
OUYANG YIHANG
TANG JUNCI
ZHAO ZHIYUAN
PAN HUAN
TANG KE
JIANG YUXIN
BAI MINGLIANG
GUO YUFENG

Assignees

国网辽宁省电力有限公司大连供电公司
国网辽宁省电力有限公司
哈尔滨工业大学

Dates

Publication Date: 20260512
Application Date: 20251224

Claims (10)

1. The probabilistic photovoltaic power prediction method based on the parameter sensitivity is characterized by comprising the following steps of: Dividing weather according to clear sky index, converting the unit and time period of short wave radiation accumulation and aligning with the measured record according to unified local time axis, processing abnormal/negative radiation reading according to solar altitude/zenith angle consistency detection rule to form a data set for modeling; The method comprises the steps of sequentially calculating direct irradiation, sky scattering and earth surface reflection components of incident surface irradiation by using the input elements, applying spectrum correction of an incident angle by adopting a cosine error correction method to obtain a relatively accurate irradiation value, combining a thermal model to obtain a state quantity of battery temperature, mapping effective irradiation and temperature into baseline power output by using an array/inverter performance model to obtain a series of deterministic baseline power prediction values, wherein the deterministic baseline power prediction values comprise component direct current power, net direct current power and grid-connected alternating current power, and form a physical chain represented by the baseline power prediction values; On the physical chain, combining the corresponding relation between the irradiation value and the baseline power predicted value, and carrying out first-order sensitivity propagation and component decomposition on uncertainty of meteorological input and device parameters; According to the state characteristics of weather/sky conditions, a light-weight data driving model of a gradient lifting tree is adopted to learn and calibrate systematic deviation and conditional variance introduced by NWP, and deterministic power prediction subjected to deviation correction and input side conditional covariance matched with the deterministic power prediction are output; and transmitting the input side conditional covariance information to a power side through a physical chain to form a multi-period conditional probability distribution and a prediction interval.
2. The method of claim 1, wherein said unit-to-period conversion of the cumulative amount of short wave radiation comprises converting the cumulative amount of short wave radiation to an average flux and performing a direct-to-scattered decomposition to yield: , ; Wherein, the Is the zenith angle of the sun, As a direct component of the ground plane, Is the total horizontal irradiance.
3. The method of claim 1, wherein a clear sky model is used to calculate a clear sky total horizontal irradiance when obtaining numerical weather forecast meteorological drives and corresponding measured photovoltaic power/irradiance data And constructing a clear sky index for weather classification K c (t): 。
4. the method of claim 1, wherein the direct, sky-scatter and surface-reflection components of the incident surface irradiance are calculated sequentially by direct irradiance of the incident surface (POA) The calculation formula of (2) is as follows: ; In the formula, Is that The normal direct irradiance at the moment; Is that Sunlight incident angle at moment; the said ground reflection irradiance The calculation formula of (2) is as follows: ; In the formula, Ground reflectivity (albedo); The installation inclination angle of the photovoltaic module is set; the total irradiance of the incident surface The three-dimensional light-emitting device consists of three parts of direct irradiation, scattering and ground reflection, and the calculation formula is as follows: ; In the formula, Is that The incident surface scattering irradiance at the moment; Finally, based on the direct irradiation and scattered irradiation decomposed into irradiation values in the vertical direction and the horizontal direction, an input element is constructed, and the specific calculation method is that the spectrum correction of the incident angle is applied to obtain the effective irradiance : ; In the formula, Correcting the coefficient for the incident angle; The coefficients are modified for spectral mismatch.
5. The method of claim 1, wherein the battery temperature The calculation formula of (2) is as follows: ; In the formula, Is the ambient air temperature; Is the wind speed; adopting Faiman temperature model; Component DC power The calculation formula of (2) is as follows: ; In the formula, Is the effective irradiance; Adopting a Mordi sub-photovoltaic array performance model; The net DC power The line loss is considered, and the calculation formula is as follows: ; In the formula, Is the loss coefficient of the direct current side; The grid-connected alternating current power The calculation formula of (2) is as follows: ; In the formula, A Sandia grid-tie inverter model was used.
6. The method of claim 1, wherein the power is treated as a deterministic map of weather and physical parameters when first order sensitivity propagation and component decomposition is performed on the uncertainty of the weather input and device parameters: ; And gives a theoretical uncertainty decomposition according to the full variance theorem: ; In the formula, The method is used for the mathematical expectation operator, And weather input vectors are used for numerical weather forecast.
7. The method of claim 1, wherein each of the gasgrams is driven based on a full probability variance decomposition Defining prediction errors and parting deviations: , ; In the formula, Is the first Weather type of individual meteorological variables The following systematic deviation expectations; Is that Original forecast error of moment; Is that Weather classification results at the moment; A mathematical expectation operator; And order Using gradient-lifted trees to characterize states Learning condition variance: ; In the formula, Conditional variance (variance estimate) for prediction; zero mean residual error after removing systematic deviation; a shape feature vector for the input model; To aim at the first The individual variables are at Gradient lifting tree regression model functions under weather-like conditions.
8. The method of claim 1, wherein the multi-period conditional probability distribution and prediction interval is subjected to consistency test and calibration by PICP, and a final probabilistic photovoltaic power prediction result is output.
9. A probabilistic photovoltaic power prediction system based on parameter sensitivity, comprising: The system comprises a data preprocessing unit, a solar elevation angle/zenith angle consistency detection rule processing unit, a solar elevation angle detection unit and a solar elevation angle detection unit, wherein the data preprocessing unit is used for acquiring numerical weather forecast meteorological drive and corresponding photovoltaic power/irradiance actual measurement data; the power resolving unit is connected with the data preprocessing unit and is used for calculating and obtaining a series of deterministic baseline power predicted values; the power estimation unit is connected with the power calculation unit and is used for learning and calibrating systematic deviation and conditional covariance introduced by the NWP and outputting deviation-corrected deterministic power prediction and matched input side conditional covariance; The output unit is connected with the power estimation unit and is used for transmitting the input side conditional covariance information to the power side through a physical chain to form a multi-period conditional probability distribution and a prediction interval; And the power supply unit is used for providing stable power supply for the three unit modules.
10. A computer-readable storage medium comprising a processor and a memory, wherein the memory contains a program for the method according to claim 1-8, and the processor is configured to process the program in steps.

Description

Probabilistic photovoltaic power prediction method and system based on parameter sensitivity Technical Field The invention relates to the technical field of power systems. Background With the urgent demands of global energy strategy for low carbonization transformation and coping with climate change, renewable energy represented by Photovoltaic (PV) power generation presents a high-speed development situation, and its permeability in power systems continues to rise. Photovoltaic power generation plays a central role in optimizing energy structures, but is strongly affected by weather conditions, exhibiting significant intermittence, randomness, and volatility. This inherent uncertainty presents a significant challenge for safe and stable operation and planning scheduling of the power system. In order to efficiently consume large-scale photovoltaic power, and simultaneously ensure the reliability and economy of a power grid, accurately quantifying uncertainty in photovoltaic power prediction has become a crucial task. Compared with the traditional single-point predicted value, the probabilistic prediction can provide a confidence interval or probability distribution of the predicted result, and has higher application value for the power grid operators to make risk perception decisions, optimize spare capacity configuration and formulate economic power market bidding strategies. Currently, one of the technical paths for solar photovoltaic power prediction is the reliance on numerical weather forecast (Numerical Weather Prediction, NWP) systems. The NWP simulates future atmospheric states through a physical model, provides key meteorological inputs such as solar irradiance and the like, and converts weather forecast into power forecast through a physical model chain (model-chain) of the photovoltaic power station. However, although NWP and physical model chain technologies are continuously mature, predictions based on this approach still face a deep technological bottleneck. First, the radiation prediction products of NWP themselves have non-negligible systematic and heteroscedastic errors. For example, the radiation field of the ERA5 and other analytical data exhibits errors related to geographical location and weather conditions, which, if not effectively corrected, can be propagated directly to the downstream photovoltaic power output, becoming a key factor in limiting the accuracy of the prediction. In addition, the physical model chain itself may introduce uncertainty due to inaccurate parameters or simplified models, resulting in problems of under-dispersion (under-dispersion) and under-calibration (miscalibration) of the predicted result, which need to be corrected by post-processing techniques. In addition to systematic deviations, the deterministic prediction information provided by the prior art prediction techniques has failed to meet the refined operating requirements of modern power grids. A single predictive value cannot quantify the inherent uncertainty in the prediction process, and the grid scheduler just needs to know the predicted risk range to formulate a more robust scheduling strategy. Thus, providing well-calibrated probabilistic predictions has become a new criterion for measuring the advancement of predictive techniques. To address the above problems, the academia and industry have explored various improvements such as post-processing of the NWP output for bias correction, or employing pure data driven probabilistic predictive models. However, these approaches often fail to deeply fuse the physical mechanism with the data-driven uncertainty quantization. For example, decoupling of the physical model and the source of uncertainty is inadequate, and it is difficult to clearly ascribe and explain the source of uncertainty. The existing research shows that the combined treatment of uncertainty from weather forecast and uncertainty from physical characteristics of a power station and the mixed strategy of merging physical models and statistical learning are effective paths for improving probability prediction skills and calibration degree. In view of the foregoing, there is a great need in the art for an innovative prediction method capable of deep fusion of physical mechanisms with data-driven models. The method not only can identify and correct time-varying deviation and error in NWP input, but also can generate probability information which is strictly calibrated and can truly reflect prediction uncertainty under different weather conditions on the basis, thereby providing comprehensive and powerful technical support for safe and economic operation of a high-proportion photovoltaic power system. Disclosure of Invention The invention aims to provide a probabilistic photovoltaic power prediction method based on parameter sensitivity, which aims to solve the limitation of the existing prediction technology in quantifying the uncertainty of photovoltaic power generation, in particular to the