CN-121998208-A - Photovoltaic power prediction method and device based on localized weather interpolation

CN121998208ACN 121998208 ACN121998208 ACN 121998208ACN-121998208-A

Abstract

The invention provides a photovoltaic power prediction method and a device based on localized weather interpolation, which relate to the technical field of photovoltaic power generation prediction, other target sites which lack weather data are enabled to learn the relation between the power of a source site with complete weather data and weather variables through a localized weather interpolation model, and the local weather interpolation is finished, missing weather data is generated, complete weather characteristic input is provided for subsequent prediction, and the problem of prediction accuracy reduction caused by missing weather data is effectively solved. The mapping relation between power and weather is established through transfer learning instead of directly performing geospatial interpolation, so that missing weather data can be generated under supervision learning, the problem that a prediction result is easy to distort when the complex terrain or weather is suddenly changed due to the assumption that the weather data is smoothly changed in space in the conventional method can be solved, and the method can still maintain higher prediction precision under the environment with severe change of the complex terrain or weather field.

Inventors

CHEN ZEFAN
ZHU JINCHENG
HU YITAO
YING ZHENHUA
HU XIANGRONG
HE XIN
LI ZHIHAO
QIAN TIANCHENG
CHEN XINBIN
Gui Yanhao
TANG YAJIE
NI CHOUWEI
Mo Hanxiang
Zhao Kaimei
CHEN ZHOUHAO

Assignees

国网浙江省电力有限公司磐安县供电公司
国网浙江省电力有限公司金华供电公司
国网浙江省电力有限公司电力科学研究院

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (12)

1. The photovoltaic power prediction method based on the localized weather interpolation is characterized by comprising the following steps: The method comprises the steps of taking a distributed photovoltaic station with complete meteorological data in an area as a source station, acquiring real meteorological data, actual measurement photovoltaic power and geographic positions of each period of the source station, and taking the rest distributed photovoltaic stations which lack complete meteorological data or lack meteorological data in the area as target stations, and acquiring actual measurement photovoltaic power and geographic positions of each period of the target stations; Based on real meteorological data, actual measured photovoltaic power and geographic positions of each period of a source station, a migration learning model is built by utilizing the relation between the photovoltaic power and meteorological variables to serve as a regional localization meteorological interpolation model, the actual measured photovoltaic power of each period of a target station is input into the regional localization meteorological interpolation model, the relation between the photovoltaic power and the meteorological variables in the learning model is obtained, and generated meteorological data corresponding to the actual measured photovoltaic power of each period of the target station is obtained; Constructing a combined data set, wherein the combined data set comprises actual measured photovoltaic power and geographic positions of all stations in the area in each period, real meteorological data of each period of a source station and generated meteorological data of each period of a target station; And (3) building a photovoltaic power prediction model, inputting the combined data set, outputting photovoltaic power prediction sequences of all the distributed photovoltaic sites in a future period of time, and obtaining a prediction result.
2. The method for predicting the photovoltaic power based on the localized weather interpolation according to claim 1, wherein the step of constructing a migration learning model as a regional localized weather interpolation model by using the relationship between the photovoltaic power and the weather variable based on the real weather data, the measured photovoltaic power and the geographical position of each period of the source site, inputting the measured photovoltaic power of each period of the target site into the regional localized weather interpolation model, learning the relationship between the photovoltaic power and the weather variable in the model, and obtaining the generated weather data corresponding to the measured photovoltaic power of each period of the target site specifically comprises the steps of: The method comprises the steps of carrying out data alignment on acquired real meteorological data, actual measured photovoltaic power and geographic position of each period of a source station and actual measured photovoltaic power and geographic position of each period of a target station, and extracting meteorological features and corresponding power features of each period of the source station and power features of each period of the target station; Defining a cross-domain mapping function taking the power characteristic as an input meteorological characteristic and taking the input meteorological characteristic as an output by utilizing the transfer learning model parameter, and mapping the power characteristic into the meteorological characteristic; Selecting meteorological features and corresponding power features of at least a part of time period of a source station as a training set for training a cross-domain mapping function to obtain a cross-domain mapping function after training; Inputting the power characteristics of each period of the target site into the cross-domain mapping function after training to predict; and predicting and generating weather features corresponding to the power features of each period of the target site, thereby obtaining the generated weather data.
3. The photovoltaic power prediction method based on localized weather interpolation according to claim 2, wherein the weather features and the corresponding power features of at least a part of the time period of the source site are selected as training sets for training the cross-domain mapping function, and a cross-domain mapping function after training is obtained, wherein in the training process, a joint loss function formed by reconstruction loss and maximum mean difference loss is designed, and the joint loss function is optimized with minimum joint loss as a target, and model parameters are iteratively updated until the model converges, so that training is completed.
4. The photovoltaic power prediction method based on localized weather interpolation according to claim 1, wherein the step of establishing a photovoltaic power prediction model, inputting a combined data set, outputting photovoltaic power prediction sequences of all distributed photovoltaic sites in a future period of time, and obtaining a prediction result comprises the steps of establishing an improved space-time diagram convolution network model based on a data source perception evaluation mechanism as a prediction model, inputting the combined data set, distinguishing real data from generated data through the data source perception evaluation mechanism, dynamically evaluating the credibility of the generated data, performing graph convolution and time sequence convolution after feature fusion based on the credibility, capturing the spatial dependency relationship and the time dependency relationship between sites, and outputting the photovoltaic power prediction sequences of all distributed photovoltaic sites in the future period of time, and obtaining the prediction result.
5. The photovoltaic power prediction method based on localized weather interpolation according to claim 4, wherein the step of establishing an improved space-time diagram convolutional network model based on a data source perception evaluation mechanism as a prediction model, inputting the merged data set, distinguishing real data from generated data through the data source perception evaluation mechanism, dynamically evaluating the credibility of the generated data, and performing feature fusion based on the credibility specifically comprises: On the basis of a space-time diagram convolution network comprising a space convolution layer, a time convolution layer and a full connection layer, constructing a space-time diagram structure comprising a weight matrix block, an explicit gating unit and a regularization constraint mechanism and placing the space-time diagram structure in front of the space convolution layer, and obtaining an improved space-time diagram convolution network model based on a data source perception evaluation mechanism as a prediction model; carrying out data alignment and feature extraction on the combined data set to obtain a combined feature set containing geographic features, power features, real meteorological features and generated meteorological features, and inputting the combined feature set into the improved space-time diagram convolutional network model; splitting the combined feature set according to sources based on an improved space-time diagram convolution network model, distinguishing real features from generated features, and respectively endowing different weight matrixes to various features of each site; And independently setting a dynamic gating mechanism for the generated meteorological features to dynamically evaluate the credibility of the generated meteorological features, dynamically adjusting a weight matrix of the generated meteorological features according to the credibility, and obtaining the meteorological-geographic-power fusion features of each site after weighted fusion.
6. The photovoltaic power prediction method based on localized weather interpolation according to claim 5, wherein the dynamic gating mechanism is specifically that the generated weather features are input into a multi-layer perceptron, the multi-layer perceptron outputs the credibility of the generated weather features according to the approaching degree of the generated weather features and the real weather features, the output of the multi-layer perceptron outputs a gating coefficient between 0 and 1 through a Sigmoid function, and a weight matrix of the generated weather features is adjusted through the gating coefficient.
7. The photovoltaic power prediction method based on localized weather interpolation according to claim 1, wherein the step of obtaining real weather data, measured photovoltaic power and geographic position of each period of the source station by using the distributed photovoltaic station with complete weather data in the area as the source station specifically comprises the step of obtaining real weather data, measured photovoltaic power and geographic position of each period of each source station by using a plurality of distributed photovoltaic stations with complete weather data in the area as a plurality of source stations.
8. The method for predicting the photovoltaic power based on the localized weather interpolation according to claim 7, wherein the step of constructing the migration learning model as the regional localized weather interpolation model by using the relationship between the photovoltaic power and the weather variable based on the real weather data, the measured photovoltaic power and the geographical position of each period of the source station specifically comprises constructing the migration learning model as a corresponding one of the regional localized weather interpolation models by using the relationship between the photovoltaic power and the weather variable based on the real weather data, the measured photovoltaic power and the geographical position of each period of each source station, wherein the number of the regional localized weather interpolation models corresponds to the number of the source stations.
9. The method for predicting the photovoltaic power based on the localized weather interpolation according to claim 8 is characterized in that the step of inputting the measured photovoltaic power of each period of the target site into a localized weather interpolation model of a region, learning the relation between the photovoltaic power and weather variables in the model, and obtaining the generated weather data corresponding to the measured photovoltaic power of each period of the target site is specifically included in the steps of inputting the measured photovoltaic power of each period of the target site into the localized weather interpolation model of each region, learning the relation between the photovoltaic power and weather variables in each model, outputting a plurality of groups of first generated weather data corresponding to the measured photovoltaic power of each period of the target site, introducing the first generated weather data outputted by the integrated gate control network to dynamically weight and integrate each localized weather interpolation model, and obtaining the second generated weather data as final generated weather data for the subsequent steps.
10. The method of claim 4, further comprising constructing a correction model based on the acquired weather forecast data for a future period of time, and correcting the output power forecast results of all the distributed photovoltaic sites for the future period of time.
11. The method for predicting photovoltaic power based on localized weather interpolation according to claim 10, wherein the step of constructing a correction model based on the weather forecast data in the future period of time and correcting the power prediction results of all the distributed photovoltaic sites in the future period of time comprises: converting the weather forecast data of each period into reference photovoltaic power to obtain a reference photovoltaic power prediction sequence; constructing a self-adaptive fusion corrector, which is used for dynamically evaluating the reliability of the photovoltaic power prediction sequence output by the improved space-time diagram convolution network model and the reference photovoltaic power prediction sequence converted by weather forecast data, carrying out weighted fusion, and carrying out training based on the reference photovoltaic power prediction sequence and the photovoltaic power prediction sequence in a history period; And (3) inputting the photovoltaic power prediction sequence and the reference photovoltaic power prediction sequence output by the improved space-time diagram convolutional network model in the current period into a self-adaptive fusion corrector after training is completed for correction, and obtaining a corrected photovoltaic power prediction sequence.
12. A photovoltaic power prediction apparatus based on localized weather interpolation for performing the method of any one of the preceding claims 1 to 11, comprising: The data acquisition module is used for taking a distributed photovoltaic station with complete meteorological data in an area as a source station, acquiring real meteorological data, actual measurement photovoltaic power and geographic positions of each period of the source station, and taking the rest distributed photovoltaic stations which lack complete meteorological data or lack meteorological data in the area as target stations, and acquiring actual measurement photovoltaic power and geographic positions of each period of the target stations; The localized weather interpolation module is used for constructing a migration learning model to be used as a regional localized weather interpolation model based on real weather data, actual measured photovoltaic power and geographic positions of each period of a source station by utilizing the relation between the photovoltaic power and weather variables, inputting the actual measured photovoltaic power of each period of a target station into the regional localized weather interpolation model, and learning the relation between the photovoltaic power and weather variables in the model to obtain generated weather data corresponding to the actual measured photovoltaic power of each period of the target station; The data merging module is used for constructing a merging data set which comprises actual measured photovoltaic power and geographic positions of all the sites in the area in each period, real meteorological data of each period of the source site and generated meteorological data of each period of the target site; and the photovoltaic power prediction module is used for establishing a photovoltaic power prediction model, inputting the combined data set, outputting photovoltaic power prediction sequences of all the distributed photovoltaic stations in a future period of time, and obtaining a prediction result.

Description

Photovoltaic power prediction method and device based on localized weather interpolation Technical Field The invention relates to the technical field of photovoltaic power generation prediction, in particular to a photovoltaic power prediction method and device based on localized weather interpolation. Background Distributed photovoltaics are an important form for promoting the rapid development of renewable energy sources by virtue of flexible layout, lower access threshold, on-site digestion and the like. The generated power has more remarkable intermittence and fluctuation under high proportion access, and the strong randomness and uncertainty easily impact the safety and stability of the power system when the grid connection is performed on a larger scale. The high-precision distributed photovoltaic power prediction research has important significance for power grid dispatching, load balancing and new energy consumption, and the distributed photovoltaic ultra-short-term power prediction is crucial to guaranteeing the real-time power supply reliability of a power grid and the economic operation of a future power market. Because of the small capacity of a single site, the current distributed photovoltaic power prediction research is mainly focused on regional prediction. The research method mainly comprises two categories, namely statistics oriented to a polymer or region integral prediction scene and a traditional machine learning and neural network model, such as an autoregressive moving average model (Autoregressive Integrated Moving Average, ARIMA), gradient lifting trees (eXtreme Gradient Boosting, XGBoost), a Long Short Term Memory network (LSTM) and the like, wherein the method can mine a certain time rule but has limited modeling capability on complex space-time correlation, and the second category is a deep learning method which is raised in recent years, is superior in capturing complex space-time correlation and gradually becomes research hotspots in the field of photovoltaic prediction, such as a convolutional neural network (Convolutional Neural Network, CNN), a graph neural network (Graph Neural Network, GNN) and a graph convolution network (Graph Convolutional Network, GCN) and a space-time graph convolutional network (Spatio-Temporal Graph Convolutional Network, STGCN) which are improved based on the method. The regional distributed photovoltaic power prediction method analyzes the space-time correlation of regional multi-site power, but the input characteristics are generally focused on time sequence power data. The low-voltage distributed photovoltaic station is not provided with a local weather monitoring device or the characteristics of forecast weather resources, so that the regional forecast lacks key objective weather characteristics, and the improvement of station forecast performance is severely restricted. The geographic and climatic differences among different distributed photovoltaic stations enable a small number of meteorological features to be incapable of representing the integral space-time coupling correlation characteristics among regional meteorological, so that the training effect of a data-driven prediction model is affected. Therefore, in the distributed photovoltaic ultra-short term power prediction, the lack of local monitoring meteorological data becomes a key bottleneck for further improving the power prediction accuracy. Considering that the prior researches are used for carrying out geographic interpolation calculation on the meteorological data of each distributed photovoltaic site, the main methods comprise an inverse distance weighting method (INVERSE DISTANCE WEIGHT, IDW), a Kriging method and the like, however, the spatial distribution of the meteorological data is not uniform and smooth, and the accuracy of the spatial interpolation method under the lack of supervised learning is obviously reduced when the complex topography or the meteorological field is severely changed. Disclosure of Invention In order to solve the problems that the existing distributed photovoltaic power prediction technology, especially the distributed photovoltaic ultra-short term power prediction technology, has insufficient prediction precision due to lack of local monitoring meteorological data and the precision of a spatial interpolation method under lack of supervised learning is obviously reduced when complex terrain or meteorological field changes severely, the invention provides a photovoltaic power prediction method and device based on localized weather interpolation, which are used for effectively solving the problem of prediction accuracy reduction caused by local weather data loss, and keeping high prediction accuracy in a complex environment. In order to achieve the above purpose, the present invention provides the following technical solutions: The photovoltaic power prediction method based on localized weather interpolation comprises the steps of taking a distributed