平衡圆周搜索的空调启动时间深度学习预测模型研究
Research on deep learning prediction model based on balanced circle search for air conditioner start-up time
摘要：
针对数据驱动建模方法只关注特征映射,没有考虑过程变量之间的长期相互依赖关系,缺乏数据之间的上下文信息,忽略了不同变量之间的重要性,从而导致预测性能欠佳问题,提出了一种结合时间卷积网络、双向门控循环单元和注意力机制的空调启动时间预测模型,对该模型的4个重要参数进行了平衡圆周搜索,以提高该模型的预测性能。采用某卷烟厂实际运行数据进行了对比实验,结果表明:与基准模型相比,圆周搜索模型和平衡圆周搜索模型的预测性能分别提高了28.98%和37.91%;对于一些异常工况,与人工凭经验得到的启动时间相比,该预测模型获得的启动时间缩短了45%左右,从而降低了空调能耗。
Abstract：
In view of the fact that the data-driven modeling methods only focus on feature mapping without concerning on the long-term interdependence between process variables, lack contextual information between data and ignore the importance of different variables, which leads to poor prediction performance, a prediction model of air conditioner start-up time combining temporal convolutional network, bidirectional gated recurrent unit and attention mechanism is proposed, and four important parameters are searched in a balanced circle to improve the prediction performance of the model. Subsequently, experiments based on a cigarette factory dataset are conducted to verify the superiority of the presented model compared to other models. The results show that compared with the benchmark model, the prediction performance of the circle search model and the balanced circle search model is improved by 28.98% and 37.91%, respectively. For some abnormal working conditions, the start-up time obtained by the presented model is shortened by about 45% compared with the start-up time obtained by manual experience, thereby reducing the energy consumption of air conditioning.
Keywords:air conditioner start-up time; temporal convolutional network; bidirectional gated recurrent unit; attention mechanism; deep learning; balanced circle search