CN-122026316-A - Photovoltaic power prediction method and device

CN122026316ACN 122026316 ACN122026316 ACN 122026316ACN-122026316-A

Abstract

The invention relates to the technical field of power generation prediction and discloses a photovoltaic power prediction method and device, wherein the method comprises the steps of obtaining historical photovoltaic data in a preset time period, and sequentially carrying out hierarchical interpolation, time alignment, hierarchical anomaly correction and dynamic normalization processing to obtain preprocessing data; the method comprises the steps of generating a training sample and a verification sample based on preprocessing data, training a plurality of types of original models by using the training sample until preset conditions are met, taking the training sample as a basic model, freezing the basic model, stacking according to pairing rules to obtain a target stacking model, screening a target prediction model from the target stacking model and the basic model based on the verification sample, and executing photovoltaic power prediction operation. The method solves the problems of inaccurate prediction caused by data loss, time dislocation, incomplete anomaly detection, distributed drift, high redundancy and correlation of model fusion, unreliable evaluation and lack of multi-step output capability in the photovoltaic power prediction in the prior art.

Inventors

ZHU HAIFENG
ZHANG HENG
LIN BINGYU
Ning Zengkun
PAN DONGHUA

Assignees

固德威技术股份有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. A method of photovoltaic power prediction, the method comprising: acquiring historical photovoltaic data in a preset time period, and sequentially performing hierarchical interpolation processing, time alignment processing, hierarchical anomaly correction processing and dynamic normalization processing on the historical photovoltaic data to obtain preprocessing data; Training samples and verification samples generated based on the pre-processing data; Respectively training different types of original models by using the training samples until the original models meet preset conditions, and taking the original models meeting the preset conditions as basic models; freezing the basic model, and stacking the basic model according to a model pairing rule to obtain a target stacking model; And screening a target prediction model for photovoltaic power prediction from the target stacking model and the basic model based on the verification sample, and executing corresponding prediction operation through the target prediction model.
2. The method of claim 1, wherein the hierarchical interpolation process comprises: identifying a start-stop time stamp covered by a missing area in the historical photovoltaic data; Acquiring effective historical data corresponding to the start-stop time stamp in a preset photovoltaic data set; And filling the missing area in the historical photovoltaic data by using the effective historical data to obtain first photovoltaic data.
3. The method according to claim 2, wherein the obtaining valid history data corresponding to the start-stop time stamp in a preset photovoltaic data set comprises: determining a first reference time period in a preset photovoltaic data set according to a first period window; inquiring whether first historical data corresponding to the start-stop time stamp exists in the first reference time period; If the first historical data exist in the first reference time period, the first historical data in the first period window are used as the effective historical data; Or if the first historical data does not exist in the first reference time period, determining a second reference time period in a preset photovoltaic data set according to a second period window, inquiring second historical data corresponding to the start-stop time stamp in the second reference time period, and taking the second historical data in the second period window as the effective historical data.
4. The method of claim 1, wherein the time alignment process comprises: acquiring first photovoltaic data obtained after hierarchical interpolation processing, and extracting irradiance sequences and photovoltaic power sequences in the first photovoltaic data; Translating the irradiance sequence within a preset sliding time offset range to generate a translated irradiance sequence; calculating a correlation coefficient of each offset between the photovoltaic power sequence and the irradiance sequence after translation, and selecting the offset with the maximum correlation coefficient as a target offset for translation of the irradiance sequence; If the target offset is not the preset value, filling the head and tail vacant positions of the irradiance sequence after translation with a preset filling value to obtain a corrected irradiance sequence, wherein the filling value comprises zero value, adjacent value or interpolation value; And taking the photovoltaic power sequence and the corrected irradiance sequence as second photovoltaic data.
5. The method of claim 1, wherein the hierarchical anomaly correction process comprises: Obtaining second photovoltaic data obtained after time alignment treatment, and calculating a power value of each sunlight volt according to the second photovoltaic data; determining an effective data period from the daily photovoltaic power value in the preset time period; grouping the second photovoltaic data in the effective data period according to a preset interval to which the irradiance value belongs by taking the irradiance level as a grouping condition to obtain a plurality of irradiance groupings; Respectively calculating box diagram statistics for second photovoltaic data in each irradiance group, identifying outliers, and correcting the outliers to obtain corrected photovoltaic data; Establishing a target regression model between photovoltaic power and irradiance based on the modified photovoltaic data; And obtaining regression residual error distribution of the target regression model, and adjusting residual error abnormal points in the corrected photovoltaic data according to the regression residual error distribution to obtain third photovoltaic data.
6. The method of claim 5, wherein the dynamic normalization process comprises: Obtaining third photovoltaic data obtained after layering abnormality correction processing; And dividing the time-varying rated power at the corresponding moment by using the third photovoltaic data to obtain fourth photovoltaic data, wherein the fourth photovoltaic data is the preprocessing data.
7. The method of claim 1, wherein stacking the base models according to model pairing rules to obtain a target stack model comprises: combining the basic models according to the model pairing rules to obtain a candidate model group; stacking the candidate model groups by using initial weight parameters to obtain an initial stacking model; And training the fusion layer in the initial stacking model by using the training sample to obtain the target stacking model.
8. The method of claim 7, wherein training the fusion layer in the initial stack model using the training samples to obtain the target stack model comprises: Acquiring a sample predicted value output by the basic model based on the training sample in the training process; inputting the sample predicted value into a fusion layer in the initial stacking model to obtain a target predicted value, wherein the fusion layer is arranged behind a candidate model group in the initial stacking model; And adjusting the weight parameters of the fusion layer in the initial stacking model by using the target predicted value and the actual value corresponding to the training sample to obtain the target stacking model.
9. The method of claim 1, wherein the screening the target prediction model for photovoltaic power prediction from the target stack model and the base model based on the validation sample comprises: extracting an input value and an actual value in the verification sample; respectively inputting the input values into the basic model and the target stacking model to obtain predicted values output by each model; calculating error indexes between the predicted value output by each model and the actual value in the verification sample; And screening the model with the minimum error index from the target stacking model and the basic model to be used as the target prediction model.
10. A photovoltaic power generation apparatus, the apparatus comprising: the acquisition module is used for acquiring historical photovoltaic data, and sequentially carrying out hierarchical interpolation processing, time alignment processing, hierarchical anomaly correction processing and dynamic normalization processing on the historical photovoltaic data to obtain preprocessing data; The generation module is used for generating training samples and verification samples based on the preprocessing data; The training module is used for training the original models of different types by utilizing the training samples respectively until the original models meet preset conditions, and taking the original models meeting the preset conditions as basic models; the stacking module is used for freezing the basic model, and stacking the basic model according to a model pairing rule to obtain a target stacking model; And the execution module is used for screening out a target prediction model for photovoltaic power prediction from the target stack model and the basic model based on the verification sample, and executing corresponding prediction operation through the target prediction model.

Description

Photovoltaic power prediction method and device Technical Field The invention relates to the technical field of power generation prediction, in particular to a photovoltaic power prediction method and device. Background With the large-scale use of renewable energy sources, the proportion of photovoltaic power generation in an electric power system is improved year by year, and the output power of the photovoltaic power generation is influenced by multiple factors such as weather, irradiance, temperature, cloud cover and the like, so that the photovoltaic power generation has remarkable volatility and certain randomness. The accurate photovoltaic power prediction is an important foundation for power system safety dispatching, energy storage control, load balancing and new energy consumption, and has a key meaning for the economical efficiency and stability of power grid operation. At present, photovoltaic power prediction is widely applied to a power dispatching center, an intelligent energy management platform and an optical storage integrated system, and is an important component of intelligent operation and maintenance of new energy. The existing photovoltaic power prediction method generally adopts a mode of combining meteorological driving with data driving, namely, historical photovoltaic power data and weather forecast data are collected and are processed through various data processing methods and then input into a single prediction model or a multi-model weighted fusion model for prediction. However, the existing method still has the defects in a plurality of key links that the existing interpolation is based on a linear or spline method, the obvious periodic characteristics of photovoltaic power generation data are not utilized, the problems of discontinuous power curve, abnormal waveforms or trend deviation and the like are easily caused under the condition of continuous deficiency or solitary point deficiency, acquisition delay or prediction deviation often exists between photovoltaic power and irradiance, if time deviation is not corrected, a model learns an incorrect input-output corresponding relation, the traditional method only adopts holiday statistics or single-point outlier detection, the distribution flow of primary screening and fine detection is not adopted, the daily level and point level abnormality are not simultaneously considered, omission or misjudgment is easily caused, the traditional method of global maximum value or fixed rated power normalization cannot be suitable for scenes such as power station expansion, overhaul or limited transmission, data distribution drift is caused, prediction errors are further caused, parameter redundancy and high correlation between models are easily generated by direct multi-model integration stacking, and accordingly fusion layer weight instability and overstock are caused, in addition, the traditional stacking method usually only focuses on longitudinal fusion of multiple models, the stacking models and single models are not trained, the structure and the model is not matched with single models, and the structure is not matched, and the time sequence characteristics are not used under different conditions, and the weather conditions are different. Although the prior art has a certain practicability in engineering, the prior art still has obvious defects in the aspects of physical rationality of data interpolation, time consistency of power and meteorological data, robustness of anomaly detection, real-time performance of rated capacity change and reliability of model fusion. Together, these problems lead to reduced stability and accuracy of photovoltaic power predictions in multi-climate, cross-season, and capacity change scenarios. Disclosure of Invention In view of the above, the embodiments of the present invention provide a method and an apparatus for predicting photovoltaic power, so as to solve the problems in the prior art that the photovoltaic power prediction is inaccurate due to data missing, time misplacement, incomplete anomaly detection, distribution drift, model fusion redundancy and correlation are high, evaluation is unreliable, and multi-step output capability is lacking. In a first aspect, an embodiment of the present invention provides a photovoltaic power prediction method, where the method includes: acquiring historical photovoltaic data in a preset time period, and sequentially performing hierarchical interpolation processing, time alignment processing, hierarchical anomaly correction processing and dynamic normalization processing on the historical photovoltaic data to obtain preprocessing data; Training samples and verification samples generated based on the pre-processing data; Respectively training different types of original models by using the training samples until the original models meet preset conditions, and taking the original models meeting the preset conditions as basic models; freezing the basic model, and