Research On The Flight Passenger Load Factor Prediction Method Based On Multi-Source Data Fusion

The demand forecasting of air passengers has always been concerned by civil aviation industry,including airline companies,ticket agents and aircraft manufacturer.For the market of aviation passenger transportation,the flight passenger load factor(FPLF)is an important indicator to measure the flight demand.Forecasting FPLF accurately will help to resolve the problems such as air tickets oversold,seat wasted,etc.It can help civil aviation practitioners to perceive market demand in advance,so as to improve the benefit management level of enterprises and provide decision support for operating departments at all levels.In previous studies,the methods of traditional time series prediction are often used to solve the problem of FPLF prediction,but the performance is not good enough.This is because the traditional methods take one flight as the research object,it only focus on the PLFs of the predicted flight in recent time,they cannot simultaneously consider the influence of many other factors on this problem.In fact,the FPLF is also affected by the PLFs of the adjacent flights.There is the departure-time correlation between the FPLFs,which refers to the fact that the PLFs of the adjacent flights on the same airline are close.At the same time,there is the departure-day correlation between the FPLFs,the FPLF also shows the temporal dependence in the temporal dimension.For example,the FPLF has the weekly pattern at the same time in the previous week and the yearly pattern at the same time in the previous year.As well,the attributes of the predicted flight and other outer factors also affect the PLFs of the predicted flight.In this thesis,we propose a novel multi-granularity time attention recurrent neural network(MTA-RNN)to consider these factors comprehensively for forecasting FPLF.This model captures temporal correlations of FPLF at different time granularities through a multi-granularity time attention mechanism.The model selects the near time interval and distant time interval FPLF data of all flights on the airline of the predicted flight according to time dependencies of different properties,then we use encoders to encoder these two-part data respectively.In each encoder,we first use LSTM unit to capture temporal dependency of different departure-times' FPLF time series belonging to same airline,then we introduce a departure-time attention mechanism to adaptively extract relevant departure times' hidden states in LSTM unit below at each departure day by referring to the previous hidden state in LSTM unit above.In the decoder,we use a departure-day attention mechanism to select relevant hidden states in LSTM unit above across all departure days.In addition,we further add the attributes of the predicted flight and outer factors,such as holidays and day of week,to predict a period of time series of PLF for the economy class of the target flight in the future.Experimental results on a real-world historical FPLF dataset demonstrate that MTA-RNN outperforms classical time series and other novel deep-learning-based prediction methods.