基于多粒度时间注意力RNN的航班客座率预测

Flight Passenger Load Factors Prediction Based on RNN Using Multi-Granularity Time Attention

准确预测航班客座率有利于处理航班机票超售、座位虚耗等问题,然而传统时间序列预测方法只关注航班近期每日客座率的变化特点,无法同时考虑其他因素的影响,预测效果不够理想。针对该问题,提出一种基于多粒度时间注意力机制的循环神经网络模型MTA-RNN。通过构建多级注意力机制获取航班客座率在不同时间粒度下的时序相关性,同时考虑航班自身属性及节假日等其他因素,得到未来一段时间内的目标航班客座率。在真实历史航班客座率数据集上的实验结果表明,MTA-RNN模型的预测准确率高于ARIMA模型、LSTM模型和Seq2seq模型。

Accurate prediction of Flight Passenger Load Factors(FPLFs)helps in addressing overbooking and overserving of the flight seats.However,traditional time series-based prediction methods only focus on variation feature of recent daily FPLFs and ignore impacts of other factors,leading to limited prediction performance.To address the problem,this paper proposes a recurrent neural network model using multi-granularity temporal attention mechanism named MTA-RNN.The model constructs a hierarchical attention mechanism to acquire the temporal correlation of FPLFs under different temporal granularities.Also,other factors including the properties of a flight,festivals and holidays are introduced into the model to compute the target FPLFs over a certain period in the future.Experimental results on datasets of real historical FPLFs show that the MTA-RNN model has a higher prediction accuracy than ARIMA,LSTM and Seq2 seq models.