• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文会议>European Photovoltaic Solar Energy Conference and Exhibition >COMPARING AND COMBINING MACHINE LEARNING AND NUMERICAL WEATHER PREDICTION MODELS FOR SOLAR FORECASTING
【24h】

COMPARING AND COMBINING MACHINE LEARNING AND NUMERICAL WEATHER PREDICTION MODELS FOR SOLAR FORECASTING

机译：比较和结合机器学习和数值天气预报模型的太阳能预测

获取原文
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

We compare the performance of Machine Learning (ML) and Numerical Weather Prediction (NWP) models for solar forecasting at 60-min steps-ahead intervals up to 72 horizons (3-days ahead), and then seek to improve the accuracy of the better performing model by combining the two models. We develop the ML forecasting model using the gradient boosting regression tree algorithm (boostTree) with Global Horizontal Irradiance (GHI) time series data for 2019 collected in Doha (Qatar). We then evaluate this ML model and the NWP model for the same year and location using the relative Root Mean Square Error (rRMSE). We conclude by developing a boostTree classifier that uses as training material the results of the evaluation of the ML and NWP models to predict the best performing forecasting model for each choice of input time series. Our results show that on average the ML model forecasts GHI with nearly an 8% lower error rate than NWP model, across the 72 one-hour steps-ahead intervals - 18.2% vs. 26% rRMSE for the ML and NWP models, respectively. On average, the combined ML-NWP model rivals the accuracy of the ML model showing a reduction in error rate (rRMSE) of up to 1.5%. We find that ML models show accurate results for day(s)-ahead forecasts, and their performance can be improved through combination with NWP models.
机译：我们比较机器学习（ML）和数字天气预报（NWP）模型的太阳能预测模型的性能，以60分钟的步骤 - 前方间隔高达72个视野（未来3天），然后寻求提高更好的准确性通过组合两个模型来执行模型。我们使用具有全球水平辐照度（GHI）时间序列数据的渐变升压回归树算法（Boosttree）在多哈（卡塔尔）收集的全球水平辐照度（GHI）时间序列数据开发ML预测模型。然后，我们使用相对根均方误差（RRMSE）来评估该ML模型和NWP模型。我们通过开发一个Boosttree分类器的培训材料来得出结论，ML和NWP模型的评估结果来预测每种输入时间序列的最佳预测模型。我们的研究结果表明，平均地为ML型号预测GHI，误差率比NWP模型近8％，分别为72小时步进间隔 - 18.2％与ML和NWP型号的RRMSE。平均而言，组合的ML-NWP模型竞争于ML模型的准确性，显示出误差率（RRMSE）的降低高达1.5％。我们发现ML型号显示最佳结果--Ahead预测，通过与NWP模型的组合可以改善它们的性能。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号