CN-121997298-A - Photovoltaic power ultra-short-term prediction method, system, equipment and medium

CN121997298ACN 121997298 ACN121997298 ACN 121997298ACN-121997298-A

Abstract

The application discloses a photovoltaic power ultra-short-term prediction method, a system, equipment and a medium, and relates to the field of photovoltaic power prediction, wherein the method comprises the steps of acquiring historical operation data of a photovoltaic power station and preprocessing; the method comprises the steps of constructing a pre-processed historical operation data into a supervision learning sample set based on a preset time window, constructing a CNN-LSTM-Attention mixed prediction model, optimizing the super parameters of the mixed prediction model by adopting a porcupine optimization algorithm to obtain an optimal super parameter combination, assigning the optimal super parameter combination to the mixed prediction model, training the mixed prediction model by utilizing the supervision learning sample set, determining a trained photovoltaic power prediction model, inputting data to be predicted into the trained photovoltaic power prediction model, and determining a photovoltaic power prediction value at a future time point.

Inventors

LI BINGLIN
Zhao Jiuzheng
LIAN YUFENG
Lan zhe
ZHANG YANAN
ZHOU JINCHENG

Assignees

长春工业大学

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. A photovoltaic power ultra-short term prediction method, comprising: Acquiring historical operation data of a photovoltaic power station, and preprocessing the historical operation data, wherein the historical operation data comprises meteorological factor data and actual photovoltaic power data corresponding to time; Based on a preset time window, constructing preprocessed historical operation data into a supervised learning sample set, taking weather factor data of a past preset historical time step in each sample in the supervised learning sample set as input characteristics, and taking actual photovoltaic power data of a future preset prediction step in each sample as output characteristics, wherein the samples are preprocessed historical operation data; The method comprises the steps of constructing a CNN-LSTM-Attention mixed prediction model, wherein the CNN-LSTM-Attention mixed prediction model comprises a CNN feature extraction layer, an LSTM time sequence modeling layer, an Attention mechanism layer and an output layer which are connected in sequence; Optimizing the superparameter of the CNN-LSTM-attribute mixed prediction model by adopting a porcupine optimization algorithm to obtain an optimal superparameter combination, and assigning the optimal superparameter combination to the CNN-LSTM-attribute mixed prediction model, wherein the porcupine optimization algorithm is that after the position of a porcupine population is updated, chaotic mapping is introduced to disturb the individual position in the porcupine population; And training the assigned CNN-LSTM-attribute mixed prediction model by using the supervision and learning sample set, determining a trained photovoltaic power prediction model, inputting data to be predicted into the trained photovoltaic power prediction model, and determining a photovoltaic power prediction value at a future time point.
2. The photovoltaic power ultra-short term prediction method according to claim 1, the method is characterized by comprising the steps of: identifying and correcting abnormal values in the historical operation data by adopting a quartile range method; smoothing the corrected historical operation data by adopting a moving average method; Based on the smoothed historical operation data, calculating a correlation coefficient between each meteorological factor data and actual photovoltaic power data, and screening out a plurality of meteorological factor data with highest correlation with the actual photovoltaic power data as key features of model input; and mapping the screened meteorological factor data and actual photovoltaic power data to a preset interval by adopting a normalization method, and determining the preprocessed historical operation data.
3. The photovoltaic power ultra-short term prediction method according to claim 2, wherein the identifying and correcting the abnormal value in the historical operating data by using a quartile range method specifically comprises: Calculating a 25 th percentile Q1 and a 75 th percentile Q3 of the historical operating data to obtain a quartile range IQR=Q3-Q1; And judging historical operation data exceeding the [ Q1-1.5IQR, Q3+1.5IQR ] interval as an abnormal value, and correcting the abnormal value by adopting a linear interpolation method.
4. The method of claim 2, wherein the key features include illumination intensity, component temperature, ambient temperature and humidity.
5. The photovoltaic power ultra-short term prediction method according to claim 1, wherein the CNN feature extraction layer comprises at least one layer of one-dimensional convolutional neural network for extracting local spatial features in input features; The LSTM time sequence modeling layer comprises a long-period and short-period memory neural network and is used for learning a long-period time sequence dependency relationship in the output characteristics of the CNN characteristic extraction layer; the Attention mechanism layer is connected behind the LSTM time sequence modeling layer and is used for carrying out weighted summation on the long-term time sequence dependency relationship output by the LSTM time sequence modeling layer; and the output layer is used for predicting and obtaining the photovoltaic power value of the future time point according to the summation result of the Attention mechanism layer.
6. The method for ultra-short-term prediction of photovoltaic power according to claim 1, wherein the ultra-parameters include learning rate, LSTM hidden unit number and L2 regularization coefficient; The introduction of the chaotic map is used for disturbing the individual positions in the porcupine population, and specifically comprises the following steps: And introducing Logistic mapping to disturb the individual positions in the porcupine population, generating new positions, comparing the new positions with the original positions, and selecting the optimal positions to enter the next iteration.
7. The photovoltaic power ultra-short term prediction method according to claim 1, wherein the constructing of the CNN-LSTM-Attention hybrid prediction model further comprises: And setting an optimization target of the CNN-LSTM-attribute mixed prediction model to minimize the root mean square error value of the CNN-LSTM-attribute mixed prediction model on a verification set.
8. A photovoltaic power ultra-short term prediction system for performing the photovoltaic power ultra-short term prediction method of any one of claims 1-7, the photovoltaic power ultra-short term prediction system comprising: the system comprises a data preprocessing module, a photovoltaic power station, a photovoltaic power generation module and a photovoltaic power generation module, wherein the data preprocessing module is used for acquiring historical operation data of the photovoltaic power station and preprocessing the historical operation data, and the historical operation data comprises meteorological factor data and actual photovoltaic power data corresponding to time; The sample construction module is connected with the data preprocessing module and is used for constructing preprocessed historical operation data into a supervision and learning sample set based on a preset time window, taking weather factor data of past preset historical time step in each sample in the supervision and learning sample set as input characteristics and taking actual photovoltaic power data of future preset prediction step in each sample as output characteristics; The model storage and calling module is used for constructing a CNN-LSTM-attribute mixed prediction model, optimizing the superparameter of the CNN-LSTM-attribute mixed prediction model by adopting a porcupine optimization algorithm to obtain an optimal superparameter combination, and assigning the optimal superparameter combination to the CNN-LSTM-attribute mixed prediction model, wherein the CNN-LSTM-attribute mixed prediction model comprises a CNN feature extraction layer, an LSTM time sequence modeling layer, an attribute mechanism layer and an output layer which are sequentially connected, and the porcupine optimization algorithm is that after the position of a porcupine population is updated, chaotic mapping is introduced to disturb the individual position in the porcupine population; The prediction execution module is respectively connected with the sample construction module and the model storage and calling module and is used for utilizing the supervised learning sample set to train the assigned CNN-LSTM-Attention mixed prediction model, determining a trained photovoltaic power prediction model, inputting data to be predicted into the trained photovoltaic power prediction model and determining a photovoltaic power predicted value at a future time point.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the photovoltaic power ultra-short term prediction method of any one of claims 1-7.
10. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program, when executed by a processor, implements the photovoltaic power ultra-short term prediction method of any of claims 1-7.

Description

Photovoltaic power ultra-short-term prediction method, system, equipment and medium Technical Field The present application relates to the field of photovoltaic power prediction, in particular to a photovoltaic power ultra-short-term prediction method systems, devices, and media. Background With the acceleration of global energy conversion, the duty cycle of photovoltaic power generation in power systems continues to climb. However, photovoltaic output is influenced by meteorological factors such as illumination intensity, temperature, humidity and the like, has obvious intermittence and volatility, and large-scale grid connection brings serious challenges to safe and stable operation of a power grid and new energy consumption. Accurate photovoltaic power prediction, especially ultra-short-term prediction of 15 minutes to 4 hours in the future, is a key technical support for real-time power grid dispatching, power station operation optimization and power market trading. At present, photovoltaic power prediction methods are mainly divided into two types: 1. Physical model-depending on high-precision numerical weather forecast (Numerical Weather Prediction, NWP) and complex photoelectric conversion equations. The method has the defects of high modeling difficulty, high calculation cost and extremely strict requirement on accuracy of weather forecast data. 2. Statistical/learning methods include conventional autoregressive integral moving average models (Autoregressive Integrated Moving Average, ARIMA), support vector regression (Support Vector Regression, SVR), etc., and recently emerging deep learning methods such as Long Short-Term Memory (LSTM), convolutional neural networks (Convolutional Neural Network, CNN), etc. The traditional method is difficult to describe the complex nonlinear relation between the meteorological factors and the power. While deep learning methods, while exhibiting greater learning ability, face two major pain points: Model structure selection is difficult because a single model (such as LSTM or CNN) cannot effectively capture spatio-temporal features at the same time. Hybrid models (e.g., CNN-LSTM) can complement one another in advantage, but their performance is highly dependent on the design of the model structure. The super-parameter optimization is difficult, the deep learning model contains a large number of super-parameters (such as learning rate, layer number, neuron number and the like), and the setting of the parameters directly affects the performance of the model. The traditional manual trial-and-error parameter adjustment is low in efficiency, unreliable in result and easy to sink into local optimum, and becomes a bottleneck for restricting further improvement of the performance of the model. Disclosure of Invention The application aims to provide a photovoltaic power ultra-short-term prediction method, a photovoltaic power ultra-short-term prediction system, photovoltaic power ultra-short-term prediction equipment and a photovoltaic power ultra-short-term prediction medium, so that the problems that model structure selection is difficult and ultra-parameter optimization is difficult can be solved. In order to achieve the above object, the present application provides the following solutions: in a first aspect, the application provides a photovoltaic power ultra-short term prediction method, which comprises the following steps. And acquiring historical operation data of the photovoltaic power station, and preprocessing the historical operation data, wherein the historical operation data comprises meteorological factor data and actual photovoltaic power data corresponding to time. Based on a preset time window, the preprocessed historical operation data is constructed into a supervised learning sample set, meteorological factor data of a past preset historical time step length in each sample in the supervised learning sample set is used as an input characteristic, actual photovoltaic power data of a future preset prediction step length in each sample is used as an output characteristic, and the samples are preprocessed historical operation data. And constructing a CNN-LSTM-Attention mixed prediction model, wherein the CNN-LSTM-Attention mixed prediction model comprises a CNN feature extraction layer, an LSTM time sequence modeling layer, an Attention mechanism layer and an output layer which are connected in sequence. And optimizing the superparameter of the CNN-LSTM-attribute mixed prediction model by adopting a coronatine optimization algorithm to obtain an optimal superparameter combination, and assigning the optimal superparameter combination to the CNN-LSTM-attribute mixed prediction model, wherein the coronatine optimization algorithm is that after the position of a coronatine population is updated, chaotic mapping is introduced to disturb the individual positions in the coronatine population to avoid the optimization algorithm from falling into local optimum. And t