西藏羊八井地区太阳短波辐照度特征及其短期预测模型对比分析

Study on the characteristics of solar shortwave irradiance and comparative analysis of short-term irradiance prediction of Yangbajing,Tibet

本研究利用西藏羊八井太阳短波辐照度观测数据分析了该地区2020—2021年的辐射时间序列分布特征,基于时间序列分析、随机森林(Random Forest,RF)和Prophet进行建模预测,通过对比研究探究三种模型在该地区的适用性以及提高模型预测精度的方法.结果表明:该地区短波太阳辐照度呈"双峰"倒"U"型分布的月变化和"单峰"倒"U"型分布的日变化特征.RF在选用模型中最优,其标准化均方根误差(Normalized Root Mean Square Error,NRMSE)、决定系数R2分别为17.54%和0.962.小波变换去噪能提高各模型预测精度,NRMSE降低4.82%~12.94%.组合模型能提高预测精度,误差倒数权重组合模型的NRMSE较差分自回归滑动平均模型(Autoregressive Integrated Moving Average,ARIMA)和Prophet分别下降35.22%、25.12%.预测时间步长差异也会影响预测效果,模型的预测误差随时间步长逐渐增大而减小.因此,可利用RF等机器学习模型在西藏地区进行太阳辐照度短期预测,通过小波变换、组合模型、预测时间步长等环节提高预测精度,以满足当地光伏发电对太阳辐照度的预测需求.

In this research,using the measured solar short-wave irradiance data during the year of 2020 and 2021 at Yangbajing observation station in Tibet,the characteristics of radiation time series distribution is analyzed.And three solar irradiance prediction models tailor to Yangbajing area are established based on time series analysis,random forest(RF)and Prophet.Moreover,by comparing this three models,the applicability of three models and the method for improving the prediction accuracy of three models are explored.Our result shows that the monthly and diurnal variation of short-wave solar irradiance in this area display a"bimodal"inverted"U"and a"unipolar"inverted"U"distribution,respectively.Among the three models,RF is found to be the best model for predicting the solar irradiance in this area,with NRMSE(Normalized Root Mean Square Error)and R2 of 17.54%and 0.962,respectively.Both wavelet transform denoising and combination model can improve the prediction accuracy of the three models and the NRMSE by applying wavelet transform denoising is reduced by 4.82%~12.94%.The NRMSE of autoregressive integrated moving average model(ARIMA)and Prophet of the error reciprocal weight combination model decreased by 35.22%and 25.12%,respectively.Furthermore,prediction time step differences also affect the prediction effect,and the prediction error of the model gradually becomes smaller with the time step.Therefore,machine learning models such as RF can be used to predict solar irradiance in Tibet,and the prediction accuracy can be improved through wavelet transforms,combined models,prediction time steps,etc.,in order to meet the forecasting needs of local photovoltaic power generation for solar irradiance.