【摘要】 小时天然气负荷预测受外部特征因素与预测方法的影响，为提高其预测精度并解决其他深度学习类模型或组合模型可解释性差、训练时间过长的问题，在引入“小时影响度”这一新特征因素的同时提出一种基于极端梯度提升树(extreme gradient boosting tress, XGBoost)模型与可解释性神经网络模型NBEATSx组合预测的方法；以XGBoost模型作为特征筛选器对特征集数据进行筛选，再将筛选降维后的数据集输入到NBEATSx中训练，提高NBEATSx的训练速度与预测精度；将负荷数据与特征数据经STL(seasonal and trend decomposition using Loess)算法分解为趋势分量、季节分量与残差分量，再分别输入到XGBoost中进行预测，减弱原始数据中的噪音影响；将优化后的NBEATSx与XGBoost模型通过方差倒数法进行组合，得出STL-XGBoost-NBEATSx组合模型的预测结果。结果表明：“小时影响度”这一新特征是小时负荷预测的重要影响因素，STL-XGBoost-NBEATSx模型训练速度有所提高，具有良好的可解释性与更高的预测准确性，模型预测结果的平均绝对百分比误差、均方误差、平均绝对误差分别比其余单一模型平均降低54.20%、63.97%、49.72%,比其余组合模型平均降低24.85%、34.39%、23.41%,模型的决定系数为0.935,能够很好地拟合观测数据。更多还原

【Abstract】 Hourly natural gas load forecasting was affected by external feature factors and forecasting methods. In order to improve the accuracy of natural gas hourly load forecasting and solve the problems of poor interpretability and long training time of other deep learning models or combination models, in this paper we introduce a new feature of "hourly influence" and propose a prediction method based on the combination of extreme gradient boosting tree(XGBoost) model and interpretable neural network model NBEATSx. The XGBoost model was used for feature screening, then the filtered and dimensionality-reduced dataset and load values were inputted into NBEATSx for training, which improves the training speed and prediction accuracy of NBEATSx. The load data and feature data were decomposed into the trend, seasonal, and residual components by the seasonal and trend decomposition using Loess(STL) algorithm, and then they were inputted into XGBoost for prediction, which reduces the influence of the noise in the original data. The two types of models mentioned above were combined by the inverse variance method to obtain the prediction results of the STL-XGBoost-NBEATSx model. The results show that the new feature of "hourly influence" is an essential factor in hourly load forecasting. The STL-XGBoost-NBEATSx not only improves the training speed, but also has good interpretability and higher prediction accuracy. The mean absolute percentage error, mean square error, and mean absolute error of the prediction results of the combined model are respectively 54.20%, 63.97%, and 49.72% lower than the rest of the single model on average, and 24.85%, 34.39%, and 23.41% lower than the rest of the combined model on average. The model has an R-squared of 0.935, which provides a good fit to the observed data.更多还原