CN-121997718-A - Deep learning irradiance numerical mode prediction correction method and system for introducing clusters

CN121997718ACN 121997718 ACN121997718 ACN 121997718ACN-121997718-A

Abstract

The invention provides a deep learning irradiance numerical mode prediction correction method and a system for introducing clusters, wherein the method is characterized in that the information sequences under the same generalized weather type are classified into the same cluster by introducing a clustering algorithm; the variation modal decomposition algorithm can decompose the historical data time series signal into various inherent modal functions, remove redundant information and obtain effective characteristics, and optimize the characteristic learning task so that the subsequent deep learning model training and error correction can be more effectively carried out.

Inventors

HU ZIWEI
QIAN KEXIANG
CHEN WEIDONG
JI ZHIXIANG
ZHANG BAOLIANG
LIU PENG
TAN JIAYING
FU JING
ZHOU LIWEI

Assignees

中国电力科学研究院有限公司

Dates

Publication Date: 20260508
Application Date: 20251231

Claims (10)

1. A deep learning irradiance numerical mode prediction correction method introducing clusters is characterized by comprising the following steps: Acquiring station observation radiation data, station observation weather elements, numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements of set time before correction time, and numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements of correction date; Clustering station observation radiation data and station observation meteorological elements by adopting a clustering algorithm, and dividing a plurality of weather types; Processing the station observation radiation data by adopting a variation modal decomposition algorithm to obtain a plurality of inherent modal functions; Based on weather types, inputting station observation weather elements, numerical weather forecast mode forecast radiation data, numerical weather forecast mode forecast weather elements and various inherent mode functions with set duration before correction time, and numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements of correction dates into a pre-trained irradiance numerical mode forecast correction model to obtain numerical weather forecast mode forecast radiation correction results of correction dates; the irradiance numerical mode prediction correction model is a CNN-LSTM network of an encoder-decoder structure, and is obtained by training historical station observation radiation data, station observation weather factors, numerical weather prediction mode prediction radiation data and numerical weather prediction mode prediction weather factors.
2. The method of claim 1, wherein the training process of the irradiance numerical model forecast correction model comprises: Acquiring station observation radiation data, station observation weather elements, numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements at historical moments, and dividing a training set and a verification set according to set proportions; Clustering station observation radiation data and station observation meteorological elements by adopting a clustering algorithm, and dividing a plurality of weather types; Processing the station observation radiation data by adopting a variation modal decomposition algorithm to obtain a plurality of inherent modal functions; Aiming at each weather type, the station observation radiation data of the correction time is taken as the CNN-LSTM network for outputting the training encoder-decoder structure, and the irradiance numerical mode forecast correction model is obtained by verifying the training set.
3. The method of claim 2, wherein the step of obtaining the irradiance numerical mode forecast correction model by using the station observation radiation data of the correction date as the CNN-LSTM network for outputting the training encoder-decoder structure and verifying with the verification set, taking as input the set of the time length station observation weather elements, the numerical weather forecast mode forecast radiation data, the numerical weather forecast mode forecast weather elements, the respective intrinsic mode functions, and the correction time, and the step of verifying with the verification set comprises: Inputting station observation meteorological elements, numerical weather forecast mode forecast radiation data, numerical weather forecast mode forecast meteorological elements and each inherent mode function of each moment in the time sequence input length of the encoder before correcting time in a training set into a CNN structure of the encoder to obtain encoder feature vectors of each moment; inputting the characteristic vector of the encoder at the current moment and the cell state and the hidden state which are output by the LSTM structure of the encoder at the last moment into the LSTM structure to obtain the cell state and the hidden state which are output by the LSTM structure at the current moment, and continuously cycling to obtain the cell state and the hidden state which are output by the encoder at the last moment before the correction date, wherein the cell state and the hidden state which are output by the encoder at the last moment at the initial moment are preset values; Inputting the numerical weather forecast mode forecast radiation data, the numerical weather forecast mode forecast meteorological elements and the correction result of the previous moment in the time sequence length of the decoder at the beginning of the correction time in the training set into the CNN structure of the decoder to obtain the decoder feature vector of each moment; inputting the characteristic vector of the decoder at the current moment and the cell state and the hidden state which are output by the LSTM structure of the decoder at the last moment into the LSTM structure to obtain the cell state and the hidden state which are output by the LSTM structure at the current moment, inputting the hidden state which is output by the LSTM structure at the current moment into a full-connection layer to obtain the correction result at the current moment, and continuously cycling to obtain the correction result at each moment in the time sequence length of the decoder at the beginning of the correction time, wherein the cell state and the hidden state which are output by the decoder at the last moment at the beginning are the cell state and the hidden state which are output by the encoder at the last moment of the correction date; And obtaining a prediction error by differencing the correction result output by the decoder and the corresponding station observation radiation data, and adjusting parameters of the encoder and the decoder based on the error until verification of a verification set is passed, so as to obtain an irradiance numerical mode prediction correction model.
4. The method of claim 1, wherein the meteorological element comprises one or more of temperature, humidity, or ground wind field component.
5. The method of claim 1, wherein the obtaining station observation radiation data, station observation weather elements, numerical weather forecast mode forecast radiation data, and numerical weather forecast mode forecast weather elements for a set duration before the correction time, and numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements for the correction date, comprises: Collecting station observation radiation data and station observation meteorological elements of a new energy station in a research area range with a set time length before correction time; Selecting a proper physical parameterization scheme of a WRF-solar mode to perform irradiance simulation experiments, and simultaneously performing dynamic downscaling experiments, and outputting grid point forecast radiation data and grid point forecast meteorological elements of set duration and correction date before correction time; And carrying out data standardization on the grid point forecast radiation data and the grid point forecast meteorological elements by adopting a batch normalization method, and interpolating the grid point forecast radiation data and the grid point forecast meteorological elements to site positions by adopting a bilinear interpolation method to obtain numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast meteorological elements matched with sites.
6. The deep learning irradiance numerical mode prediction correction system introducing the clustering is characterized by comprising a data acquisition module, a type dividing module, a mode decomposing module and a radiation correction module; the data acquisition module is used for acquiring station observation radiation data, station observation weather elements, numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements of set time before correction time, and numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements of correction date; The type dividing module is used for clustering the station observation radiation data and the station observation meteorological elements by adopting a clustering algorithm and dividing a plurality of weather types; the modal decomposition module is used for processing the station observation radiation data by adopting a variation modal decomposition algorithm to obtain a plurality of inherent modal functions; The radiation correction module is used for inputting station observation weather elements, numerical weather forecast mode forecast radiation data, numerical weather forecast mode forecast weather elements and all inherent mode functions with set duration before correction time, and numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements of correction date into a pre-trained irradiance numerical mode forecast correction model based on weather types to obtain a numerical weather forecast mode forecast radiation correction result of correction date; the irradiance numerical mode prediction correction model is a CNN-LSTM network of an encoder-decoder structure, and is obtained by training historical station observation radiation data, station observation weather factors, numerical weather prediction mode prediction radiation data and numerical weather prediction mode prediction weather factors.
7. The system of claim 6, wherein the training process of the irradiance numerical model predictive correction model in the irradiance correction module comprises: Acquiring station observation radiation data, station observation weather elements, numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements at historical moments, and dividing a training set and a verification set according to set proportions; Clustering station observation radiation data and station observation meteorological elements by adopting a clustering algorithm, and dividing a plurality of weather types; Processing the station observation radiation data by adopting a variation modal decomposition algorithm to obtain a plurality of inherent modal functions; Aiming at each weather type, the station observation radiation data of the correction time is taken as the CNN-LSTM network for outputting the training encoder-decoder structure, and the irradiance numerical mode forecast correction model is obtained by verifying the training set.
8. The system of claim 7, wherein the radiation correction module takes station observation weather elements, numerical weather forecast mode forecast radiation data, numerical weather forecast mode forecast weather elements, each inherent mode function and numerical weather forecast mode forecast radiation data of correction time as input, takes station observation radiation data of correction date as CNN-LSTM network for outputting training encoder-decoder structure, and verifies with verification set to obtain irradiance numerical mode forecast correction model, comprising: Inputting station observation meteorological elements, numerical weather forecast mode forecast radiation data, numerical weather forecast mode forecast meteorological elements and each inherent mode function of each moment in the time sequence input length of the encoder before correcting time in a training set into a CNN structure of the encoder to obtain encoder feature vectors of each moment; inputting the characteristic vector of the encoder at the current moment and the cell state and the hidden state which are output by the LSTM structure of the encoder at the last moment into the LSTM structure to obtain the cell state and the hidden state which are output by the LSTM structure at the current moment, and continuously cycling to obtain the cell state and the hidden state which are output by the encoder at the last moment before the correction date, wherein the cell state and the hidden state which are output by the encoder at the last moment at the initial moment are preset values; Inputting the numerical weather forecast mode forecast radiation data, the numerical weather forecast mode forecast meteorological elements and the correction result of the previous moment in the time sequence length of the decoder at the beginning of the correction time in the training set into the CNN structure of the decoder to obtain the decoder feature vector of each moment; inputting the characteristic vector of the decoder at the current moment and the cell state and the hidden state which are output by the LSTM structure of the decoder at the last moment into the LSTM structure to obtain the cell state and the hidden state which are output by the LSTM structure at the current moment, inputting the hidden state which is output by the LSTM structure at the current moment into a full-connection layer to obtain the correction result at the current moment, and continuously cycling to obtain the correction result at each moment in the time sequence length of the decoder at the beginning of the correction time, wherein the cell state and the hidden state which are output by the decoder at the last moment at the beginning are the cell state and the hidden state which are output by the encoder at the last moment of the correction date; And obtaining a prediction error by differencing the correction result output by the decoder and the corresponding station observation radiation data, and adjusting parameters of the encoder and the decoder based on the error until verification of a verification set is passed, so as to obtain an irradiance numerical mode prediction correction model.
9. The system of claim 6, wherein the meteorological element comprises one or more of temperature, humidity, or ground wind field component.
10. The system of claim 6, wherein the data acquisition module is specifically configured to: Collecting station observation radiation data and station observation meteorological elements of a new energy station in a research area range with a set time length before correction time; Selecting a proper physical parameterization scheme of a WRF-solar mode to perform irradiance simulation experiments, and simultaneously performing dynamic downscaling experiments, and outputting grid point forecast radiation data and grid point forecast meteorological elements of set duration and correction date before correction time; And carrying out data standardization on the grid point forecast radiation data and the grid point forecast meteorological elements by adopting a batch normalization method, and interpolating the grid point forecast radiation data and the grid point forecast meteorological elements to site positions by adopting a bilinear interpolation method to obtain numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast meteorological elements matched with sites.

Description

Deep learning irradiance numerical mode prediction correction method and system for introducing clusters Technical Field The invention relates to the technical field of irradiance correction, in particular to a deep learning irradiance numerical mode prediction correction method and system for introducing clusters. Background Currently, non-fossil energy is greatly developed worldwide, wind power generation and solar power generation are accelerated to be developed, and large-scale development and high-quality development of wind power generation and solar power generation are comprehensively promoted. In recent years, new energy power generation represented by wind power and photovoltaic has been developed in a blowout manner. By 2023, taking china as an example, the integrated installed capacity of photovoltaic power generation has reached about 610GW, and has been the second largest power form of installed capacity nationwide. However, as the installed capacity of the photovoltaic power generation increases, the photovoltaic power generation power is affected by factors such as irradiance and the like to change in volatility and randomness, so that the solar energy utilization rate is further reduced. In order to reduce the adverse effect of the instability of the photovoltaic power generation on the power grid and the waste of solar energy resources, the adoption of a photovoltaic power generation power prediction technology is an effective solution. The accurate prediction of the photovoltaic output is beneficial to the advanced planning of power dispatching, and the output of the station is reasonably arranged, so that the safety of a power grid is ensured, the electric energy quality is improved, and the stable and reliable operation of a power system is ensured. In addition, china is actively pushing electricity market, and electric power spot transaction has been realized in some areas. With further development of the power market mechanism, the ultra-short-term photovoltaic generation power prediction technology becomes more important, and regulation and control of the power market quotation strategy can be promoted and economic benefit maximization can be achieved. Currently, combining NWP numerical prediction data as input to a predictive correction model is a method that is widely used by students. The NWP solar radiation is corrected, so that the prediction result of photovoltaic power can be effectively improved. At present, compared with the traditional methods of data assimilation, model output statistics and the like, a plurality of scholars use machine learning and deep learning methods to correct prediction deviation. The deep learning model has strong characteristic learning capability, can learn higher-level abstract features from large-scale data, and is beneficial to extracting potential information and rules in meteorological data, so that numerical mode deviation correction is better carried out. In order to make up for the defects of a single correction method, many students combine the advantages of different correction methods, such as a statistical method and a deep learning correction method in a data preprocessing stage, so as to improve the accuracy of irradiance correction. And the time-space correlation of training data is enhanced while the hidden time-space characteristics in the mode forecast data are mined, and the integral deviation correcting capability of the network structure is improved, so that the forecast accuracy after correction is improved. However, the correction result is sensitive to the change of the weather condition, and most of the existing NWP GHI correction models do not consider to model separately for different weather conditions, but NWP GHI errors can have great differences due to the different weather conditions. Meanwhile, discrete variables exist in weather factors and irradiance, nonlinear mapping relations exist between the weather factors and the irradiance, and the current correction method is insufficient in terms of the relation between random and discrete processing variables. Disclosure of Invention In order to solve the problem of insufficient performance in terms of the relation between random and discrete processing variables in the prior art, the invention provides a deep learning irradiance numerical mode prediction correction method for introducing clusters, which comprises the following steps: Acquiring station observation radiation data, station observation weather elements, numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements of set time before correction time, and numerical weather forecast mode forecast radiation data and numerical weather forecast mode forecast weather elements of correction date; Clustering station observation radiation data and station observation meteorological elements by adopting a clustering algorithm, and dividing a plurality of weathe