基于XGBoost-神经网络的建筑负荷预测模型构建

Construction of Building Load Prediction Model Based on XGBoost-Neural Network Algorithm

针对建筑负荷预测模型特征选择工作量大、泛化能力提升难的问题,提出一种基于极限梯度提升(extreme gradient boosting,XGBoost)-神经网络的建筑负荷特征筛选及预测方法,利用XGBoost算法训练滤波处理后的数据,基于平均绝对误差百分比(mean absolute percentage error,MAPE)确定最优特征子集,以改善模型精度和泛化能力;采用贝叶斯正则化算法训练前馈神经网络,以便能够在训练优化过程中降低网络结构复杂性,从而避免网络过拟合,进一步提升其泛化能力。针对某商业建筑的负荷预测实验结果表明,特征筛选后较筛选前模型均方误差(mean squared error,MSE)降低43.29%,有效提高了模型预测精度;分别以贝叶斯正则化和Levenberg-Marquardt(L-M)算法对神经网络进行训练,前者5次试验均方根误差(root mean squared error,RMSE)和MAPE平均值较后者分别降低87.08%、85.33%,预测模型泛化能力得到有效提升。

In response to the problem of heavy workload and difficulty in improving the generalization ability of feature selection in building load prediction models,a method based on extreme gradient boosting(XGBoost)-neural network for building load feature selection and prediction was proposed.The XGBoost algorithm was used to train the filtered data,and the optimal feature subset was determined based on the mean absolute percentage error(MAPE) to improve the model accuracy and generalization ability.The Bayesian regularization algorithm was used to train the feedforward neural network to reduce network structure complexity during training optimization and prevent network overfitting,thereby further improving its generalization ability.Experimental results of load prediction for a commercial building show that the mean squared error(MSE) of the model is reduced by 43.29% after feature selection compared to before,effectively improving the model prediction accuracy.The neural network is trained using both Bayesian regularization and Levenberg-Marquardt(L-M) algorithms,with the former achieving an average reduction of 87.08% and 85.33% in root mean squared error(RMSE) and MAPE after 5 experiments,respectively,leading to effective improvement of the prediction model's generalization ability.