基于模态分解和机器学习的水平辐照度预测方法研究

Research on Horizontal Irradiance Prediction Method Based on Modal Decomposition and Machine Learning

摘要: 光伏电站日渐发展成为新能源的焦点，准确、可靠的水平辐照度(GHI)预测是解决光伏电站相关问题的方法之一。因此，提出了一种有效的混合模型CEEMDAN-RF-LGBM-XGB，用于预测每小时的水平辐照度。首先，完全自适应噪声集合经验模态分解(CEEMDAN)将非线性和非平稳性的气象变量序列分解为若干个模态函数，依据对水平辐照度的影响程度，随机森林将分解的模态函数进行特征提取，降低数据复杂度，达到降维的目的，完全自适应噪声集合经验模态分解和随机森林相结合得到一组简单且信息量丰富的影响因子。其次，利用轻量级梯度提升机算法对中国三个不同气候地区的水平辐照度进行预测，得到初始的水平辐照度预测值。针对预测模型训练中产生的固有误差，引入极致梯度提升算法进行误差修正，提升模型的预测性能。针对一些混合模型和独立模型对所提模型的性能进行了测试。测试结果显示，最小%RMSE (%MAE)提高了23.14% (24.45%)。实验结果证明所提出的模型具有最高的预测精度，且提出的混合结构模型均提高了单一模型的预测能力。

Abstract: Photovoltaic power plants are increasingly becoming the focus of new energy, and accurate and reliable horizontal irradiance (GHI) prediction is one of the methods to solve the problems related to photovoltaic power plants. Therefore, an effective hybrid model CEEMDAN-RF-LGBM-XGB was proposed for predicting hourly GHI. Firstly, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) decomposes the nonlinear and non-stationary meteorological variable sequences into several modal functions. According to the degree of influence on horizontal irradiance, the random forest extracts the features of the decomposed modal functions to reduce the data complexity and achieve the goal of dimensionality reduction. The combination of the complete ensemble empirical mode decomposition with adaptive noise and the random forest obtains a set of simple and informative influence factors. Secondly, the lightweight gradient boosting algorithm was used to predict the horizontal irradiance in three different climate regions of China, and the initial horizontal irradiance predicted value was obtained. Aiming at the inherent errors generated during the training of prediction models, the extreme gradient boosting algorithm is introduced for error correction to improve the predictive performance of the model. The performance of the proposed model was tested on some mixed and independent models. The test results show that the minimum %RMSE (%MAE) of the proposed model is increased by 23.14% (24.45%). The experimental results show that the proposed models have the highest prediction accuracy, and the proposed hybrid structure models improve the prediction ability of a single model.