Research On Solar Irradiance Prediction Method Based On Deep Learning Theory

In recent years,with the increasing environmental problems,the growing energy demand and the maturity of photovoltaic power generation technology,the photovoltaic power generation industry has been developing rapidly.China has also issued relevant policies to support the development of photovoltaic-related industries,and also clearly pointed out that in2030 to achieve "carbon peak",2060 to achieve "carbon neutral".In this context,vigorously developing solar energy is one of the most effective measures to achieve the "double carbon" goal.The abundance of solar resources and the achievability of photovoltaic power generation technology provide the conditions for developing solar energy,but solar energy is volatile and unstable,which creates obstacles to the regulation and safety of power systems.Accurate prediction of PV power helps the power system scheduling department to coordinate the grid connection plan and reduce the shock risk of PV grid connection.Solar irradiance is the most direct and important factor influencing PV power generation.Accurate prediction of solar irradiance has a profound impact on PV power output,and is also of great significance to the development of the entire PV industry and new energy sources.In the existing traditional prediction methods,there are problems of low prediction efficiency and low prediction accuracy.In order to improve the accuracy of solar irradiance prediction,two solar irradiance prediction models are proposed in this paper for research.The first is a solar irradiance prediction model based on discrete wavelet transform and genetic optimized deep neural network(GA-DNN)is proposed to analyze and predict the predicted data.The model uses genetic algorithm to optimize the depth neural network to improve the prediction performance of the model.Second,to further enhance the prediction model’s ability to identify and analyze weather and other influencing factors in the data,a long and short-term memory network model(CNN-A-LSTM)based on convolutional neural network incorporating attention mechanism with similar day analysis is proposed to effectively improve the prediction accuracy.This method explores the influencing factors of solar irradiance such as temperature,humidity,and air pressure by analyzing the data set and the characteristics of solar irradiance,and cleans the experimental data using normalization and outlier processing to reduce the influence of abnormal data.The deep learning algorithm based on long-and short-term memory networks can learn both short-term data and long-term data change trends,which effectively solves the problem of increasing prediction errors when sudden data changes occur in traditional models.Combined with the spatiotemporal characteristics of convolutional neural network processing data,it is more conducive to model training.The CNN-A-LSTM hybrid prediction model takes into account the spatiotemporal characteristics of the data while maintaining the prediction performance,which makes the hybrid model more adaptable to achieve more accurate prediction of solar irradiance.Based on the theoretical study in this paper,actual engineering data are used as arithmetic examples to verify the effectiveness of the proposed model.By comparing the proposed prediction model with several sets of comparative prediction models,the CNN-A-LSTM hybrid prediction model maintains a good prediction performance under ideal and non-ideal weather conditions,and the performance indicators show that the model has a high prediction accuracy.This is beneficial to the grid dispatching department for better advance planning of PV energy dispatching,which is important to maintain the security and stability of the grid.