基于贝叶斯优化ARIMA-CNN-GRU深度算法的楼宇负荷预测研究

Research on Building Load Forecasting Based on ARIMA-CNN-GRU Neural Network and Bayesian Optimization

楼宇能耗预测问题对降低能量消耗与实现合理功能至关重要。为解决楼宇能耗复杂多变的问题,采用ARIMA模型求解能耗曲线非线性部分,再通过CNN-GRU深度学习模型拟合非线性残差,并且采用贝叶斯优化算法对深度模型进行超参优化。基于不同楼宇负荷曲线的预测结果证明,贝叶斯优化算法能够提升模型精度3倍以上,所提出的ARIMA-CNN-GRU算法针对不同类型的楼宇负荷曲线预测的最大误差控制在7%以内,比通过CNN-GRU网络直接预测楼宇负载曲线精度提升2倍,能够满足不同楼宇负荷的预测。

Building energy consumption prediction is essential for energy consumption reduction and function rationalization.To solve the complex and changeable problems of building energy consumption,ARIMA model is used to solve the nonlinear part of the energy consumption curve,then CNN-GRU depth learning model is used to fit the nonlinear residuals,and Bayesian optimization algorithm is adopted for hyperparametric optimization.The prediction results based on load curves of different buildings prove that the Bayesian optimization algorithm can improve the accuracy of the model by more than 3 times,the maximum error of ARIMA-CNN-GRU algorithm proposed for different types of building load curve prediction is controlled within 7%.The accuracy value is 2 times higher than that of building load curve direct predicted through the CNN-GRU network,meeting the prediction of different building loads.