CN-121983972-A - New plateau station photovoltaic power prediction method and system based on transfer learning

CN121983972ACN 121983972 ACN121983972 ACN 121983972ACN-121983972-A

Abstract

The photovoltaic power prediction method and system for the newly built plateau station based on transfer learning comprise the steps of determining a photovoltaic station with rich historical data matched with the photovoltaic station according to altitude, solar radiation characteristics, weather conditions and hardware configuration factors of the newly built plateau station, constructing a power prediction model, wherein the prediction model comprises a characteristic extraction module and a task adaptation module, the characteristic extraction module comprises a plurality of layers of convolution layers and characteristic attention modules which are sequentially connected, the task adaptation module comprises at least 2 layers of convolution layers and full connection layers which are sequentially connected, the photovoltaic station data with rich historical data are utilized to pretrain the power prediction model, the task adaptation module in the power prediction model is utilized to retrain the historical data of the newly built plateau station, the final power prediction model after training is obtained, and photovoltaic power prediction of the newly built plateau station is carried out.

Inventors

ZHAO XINYU
SHI JIYANG
SUN CHAO
Duan Jinggua
ZHOU JIN
ZHENG XIAOQING
ZHAO LINA
ZHANG SHAOCHEN
XUE CONGCONG
ZHAO PENG
ZHANG YING
GU XIAODONG
YAN XIAOWEI
GUAN LEI
WANG KAIHUAI
LI CHENGYIN
CUI XIAOYUE
ZHANG QIANG

Assignees

中国铁路设计集团有限公司

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (8)

1. The new plateau station photovoltaic power prediction method based on transfer learning is characterized by comprising the following steps of: S1, acquiring photovoltaic power at a historical moment of a newly built plateau station to be predicted, irradiance at the historical moment, clear sky irradiance at the historical moment, temperature at the historical moment, altitude and clear sky irradiance at a future moment to be predicted; s2, constructing a power prediction model, wherein the prediction model comprises a feature extraction module and a task adaptation module, the feature extraction module comprises a plurality of layers of convolution layers and a feature attention module which are sequentially connected, the task adaptation module comprises at least 2 layers of convolution layers and a full connection layer which are sequentially connected, the output of the last layer of convolution layers in the feature extraction module is connected with the feature attention module, attention weight is obtained through the feature attention module, and the obtained attention weight and the output of the last layer of convolution layers in the feature extraction module are multiplied element by element to obtain a new matrix to serve as the input of the convolution layers in the task adaptation module; s3, pre-training the power prediction model by using photovoltaic station data with rich historical data, and retraining a task adaptation module in the power prediction model by using the historical data of the newly built plateau station to obtain a trained final power prediction model; s4, carrying out photovoltaic power prediction of the newly built plateau station by using the trained final power prediction model; the method for determining the photovoltaic field station with rich historical data matched with the photovoltaic field station comprises the following steps of: The altitude is used as a core dominant factor, the altitude difference between a source domain and a target domain is used as a quantification standard, the atmospheric physical characteristics of the source domain and the target domain are ensured to be consistent, the similarity of radiation transmission and temperature environment is fundamentally ensured, and meanwhile, the total radiation deviation of a horizontal plane is less than or equal to 15 percent and the direct radiation proportion deviation is less than or equal to 10 percent; the climate conditions take annual average relative humidity deviation less than or equal to 12 percent and annual average sand and dust daily number deviation less than or equal to 8 percent as matching indexes, so that domain distribution mismatch caused by attenuation mechanism difference is avoided; the efficiency deviation of the hardware configuration control assembly is less than or equal to 3%, the ratio of installed capacity is in a range of 0.5-2, the topological structure of the system is consistent, and the interference of equipment difference on migration effectiveness is reduced.
2. The photovoltaic power prediction method for a newly built plateau station based on transfer learning according to claim 1, wherein when determining a photovoltaic station with abundant historical data matched with the newly built plateau station according to altitude, solar radiation characteristics, climate conditions and hardware configuration factors of the newly built plateau station through multi-factor quantitative similarity sorting, the method further comprises: The method comprises the steps of firstly, taking altitude as a main control factor, executing rigid screening, only keeping the built stations with the altitude difference less than or equal to 500m from the target station to enter a candidate set, adopting a relative difference mapping method to convert the initial differences of the candidate stations passing through the altitude screening and three factors including solar radiation, climate conditions and hardware configuration of the target station into the similarity of a unified interval, and adopting equal-weight arithmetic average to obtain a comprehensive similarity score so as to finish the sequencing of candidate source domains.
3. The photovoltaic power prediction method of a newly built plateau station based on transfer learning according to claim 1, wherein the method for pre-training a power prediction model by using photovoltaic station data with rich historical data and retraining a task adaptation module in the power prediction model by using the historical data of the newly built plateau station is as follows: The method comprises the steps of firstly, pre-training a model by using a photovoltaic station with rich historical data to obtain a pre-trained model, then, performing parameter fine adjustment on the pre-trained model by using a small amount of historical data on a newly built station, freezing a feature extraction module of the pre-trained model in the fine adjustment process, updating a convolution kernel and a deviation item in the task adaptation module by error back propagation through a task adaptation module of the fine adjustment model, and performing retraining on the pre-trained model by using a small amount of historical data on the newly built station to obtain a fine-adjusted power prediction model.
4. The new plateau station photovoltaic power prediction method based on transfer learning according to claim 1, wherein the calculation formula of the convolution layer in the feature extraction module is as follows: ; Wherein, the Is an input feature of the i-th layer, Is the output characteristic of the j-th layer, Is a matrix of weights that are to be used, Is a bias matrix, symbols Is a convolution operation and is performed by, Representing a nonlinear activation function ReLU function.
5. The method for predicting the photovoltaic power of the newly built plateau station based on the transfer learning according to claim 1, wherein the method for predicting the photovoltaic power of the newly built plateau station by using the trained final power prediction model is as follows: And inputting the historical irradiance, the historical clear sky irradiance, the historical temperature, the historical photovoltaic power, the altitude at the predicted time and the clear sky irradiance at the predicted time of the newly built plateau station into a trained final power prediction model, and outputting the network of the final power prediction model as the photovoltaic power at the predicted time.
6. A newly built plateau station photovoltaic power prediction system based on transfer learning, for implementing the method of any one of claims 1 to 5, comprising: The system comprises a historical data and photovoltaic station acquisition module, a solar energy storage module and a power generation module, wherein the historical data and photovoltaic station acquisition module is used for acquiring photovoltaic power at a historical moment of a newly-built plateau station to be predicted, irradiance at the historical moment, clear sky irradiance at the historical moment, temperature at the historical moment, altitude and clear sky irradiance at a future moment to be predicted; The system comprises a power prediction model construction module, a task adaptation module, a feature attention module, a feature extraction module and a power prediction module, wherein the power prediction model construction module is used for constructing a power prediction model, the feature extraction module comprises a plurality of layers of convolution layers and a feature attention module which are sequentially connected, the task adaptation module comprises at least 2 layers of convolution layers and a full connection layer which are sequentially connected, the output of the last layer of convolution layers in the feature extraction module is connected with the feature attention module, attention weight is obtained through the feature attention module, and the obtained attention weight is multiplied by the output of the last layer of convolution layers in the feature extraction module element by element to obtain a new matrix which is used as the input of the convolution layers in the task adaptation module; The power prediction model training module is used for pre-training the power prediction model by utilizing the photovoltaic station data with rich historical data, and retraining the task adaptation module in the power prediction model by utilizing the historical data of the newly built plateau station to obtain a trained final power prediction model; and the photovoltaic power prediction module is used for predicting the photovoltaic power of the newly built plateau station by using the trained final power prediction model.
7. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the new plateau photovoltaic power prediction method based on transfer learning as claimed in any one of claims 1 to 5.
8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the new plateau photovoltaic power prediction method based on transfer learning as claimed in any one of claims 1 to 5 when executing the program.

Description

New plateau station photovoltaic power prediction method and system based on transfer learning Technical Field The invention belongs to the technical field of photovoltaic power prediction, and particularly relates to a photovoltaic power prediction method and system for a newly built plateau station based on transfer learning. Background At present, the great development of new energy sources such as solar energy has become a trend. However, photovoltaic power generation has stronger volatility and uncertainty, so that the development of high-precision photovoltaic power forecast is an effective means for promoting photovoltaic digestion and has extremely important significance. The climate environment of the plateau area is severe and complex, the installation and construction difficulties of the photovoltaic panels are high, and the difficulty of developing accurate power forecast is high for the photovoltaic stations newly operated in the plateau area with sparse historical power generation data. The photovoltaic power generation power is mainly determined by solar radiation and is also affected by temperature. In general, the higher the solar radiation degree is, the higher the generated power is, and the lower the temperature is, the stronger the power generation capability of the photovoltaic panel is. In high-altitude areas such as Qinghai-Tibet plateau, the air is thin, the air temperature is low, the conditions are bad, and the installation and construction difficulties of the photovoltaic panel are high. However, the Qinghai-Tibet plateau is one of the regions with the most abundant solar energy resources in China, and has great photovoltaic power generation potential, so that special research on photovoltaic power forecast in the plateau region is necessary. The main difference between the plateau region and other regions is the comprehensive environmental difference caused by the altitude. The high altitude results in thin air, thinned atmosphere to reduce solar radiation attenuation, obviously enhanced irradiance, low air temperature and high photovoltaic cell converting efficiency. These unique geographic and climatic conditions make the plateau region not only face Shi Gongyun-dimensional challenges in terms of photovoltaic power generation, but also have unique resource advantages; In the literature He Y, Gao Q, Jin Y, et al. Short-term photovoltaic power forecasting method based on convolutional neural network[J]. Energy Reports, 2022, 8: 54-62. , historical photovoltaic power data of a station to be forecasted is utilized to forecast future photovoltaic power, and good forecasting accuracy is displayed in a scene that the historical photovoltaic power data are sufficient. However, the existing photovoltaic power forecasting method only conducts photovoltaic power forecasting on sites with sufficient historical power data, relies on a large amount of historical power generation data, and cannot conduct high-precision photovoltaic power forecasting on new sites with new operation and sparse data. Meanwhile, in the prior art, power forecasting is often carried out on photovoltaic stations in plain or areas with low altitudes, characteristics such as altitude are not considered, accurate forecasting precision is difficult to obtain in photovoltaic forecasting in areas with extremely high altitudes such as Qinghai-Tibet plateau, and accurate photovoltaic power forecasting is more difficult to achieve in new stations in plateau areas with scarce data. Disclosure of Invention In order to solve the problem of power forecasting on a photovoltaic station with sparse data in a plateau region, the invention aims to construct a photovoltaic power forecasting method of the data-sparse station based on depth CNN (Convolutional Neural Network ) transfer learning, and the high-precision photovoltaic power forecasting of the data-sparse station in the plateau region is realized by transferring knowledge and sharing information between the photovoltaic station with rich historical data and a new station with sparse historical data and integrating the knowledge and the information into the characteristics of altitude, temperature and the like which can reflect the special natural conditions in the plateau region; For the specificity of the plateau region, when a photovoltaic power forecasting model is constructed, characteristic variables which can fully reflect the physical mechanism of the photovoltaic power forecasting model need to be selected. Besides historical photovoltaic power data, the photovoltaic power system has four characteristics of irradiance, clear sky irradiance, temperature and altitude, which form a complete physical driving prediction framework, irradiance serves as a core input variable to directly quantify energy sources of a photovoltaic system, clear sky irradiance provides a reference standard of irradiance in an ideal state without cloud layer shielding, the maximum possible value