A new forecasting system for high-speed railway passenger demand based on residual component disposing

Highlights

• A new forecasting system is presented based on residual component disposing.

• Moving block bootstrap is used to simulate residual component for future data.

• The residual is included in forecasting procedure by extrapolation principle.

• Different forecasting models are built for several subseries.

Abstract

Accurate passenger demand forecasting of high-speed railway is of great significance for railway line planning and daily operation management. Capturing the random factors hidden in complex data is the key to achieve accurate forecasting. In view of this, this paper presents a new system with deterministic and probabilistic forecasting capacities based on the residual component disposing. First, the passenger demand is decomposed into trend, seasonality components and remainder using seasonal and trend decomposition using loess. Most likely some weak but indispensable information may still be contained in the remainder component, and random incidents occurred in the past may also occur in the future. Thus, moving block bootstrap is employed to bootstrap the remainder and generate one thousand similar samples by virtue of extrapolating the disaggregated subseries, thereby simulating the stochastic feature of future series. The forecasts of similar sample are acquired through reaggregating the results of extrapolations. Eventually, the bagging predictors is applied to attain the deterministic and probabilistic prediction. Two real-world case study manifests that the proposed hybrid system provides a more accurate assessment of passenger demand than all the benchmark models in various aspects, and the theoretical framework established in this paper is of certain enlightening significance for dealing with complex data.

Introduction

Over the past decade, the extraordinary tempo of high-speed railway (HSR) development in several countries has attracted the continuous attention of the world. Especially in China, as a rapid inter-city transportation mode, HSR has reduced the heavy burdens of passenger transport for busy trunk lines among metropolises significantly [1]. Undeniably, there is a broad consensus that the accurate prediction of high-speed railway passenger flow is a significant issue for both construction and operation of the HSR system, such as planning and development of railway network, train operation plan, train or vehicle dispatching, and power supply operation management [2]. Aiming to develop relevant planning and control the system in a proactive manner, a predictive procedure has to be incorporated in the transportation system [3].

Many relevant items would be taken into consideration when the majority of travelers select trip modes, such as motivation travel, economic income, and personal preference. The passenger demand of HSR is usually affected by various factors [4]. Especially, in China, as the country with the largest population and the fastest development of high-speed railway, its high-speed railway passenger flow also has its own characteristics. Usually, the daily passenger demand during the weekend is generally greater than that of weekdays of the week, and the number of passengers during holidays and festivals outnumbers that of workday [5]. Thus the periodic behavior of travel characteristic would lead to the seasonal evolution hidden in passenger demand time series. From the perspective of development and evolution trend, in competition with aviation, common speed railway, and road infrastructure, HSR is favored by increasing travelers for a more comfortable and safer journey experience, which is the internal cause of the trend characteristics of HSR passenger flow. Both trend and seasonality are the main features of passenger demand data and are relatively easy to be captured by predictors. In the majority of cases, however, the uncertainty of individual trip mode is the main reason of chaotic and random presented in passenger demand data, which influences the forecasting procedure and results greatly. Thus, the chaotic and random property cannot be overlooked for accurate forecasting.

In most of the research work, the core work to achieve accurate prediction is how to deal with the complex characteristics of time series, such as random, chaotic, fluctuation and so on. To minimize the difficulty of prediction, a variety of data processing mechanisms have been developed to reduce the complexity of data and improve the prediction performance. In recent years, a kind of empirical mode decomposition (EMD) based on decomposition ensemble technology can effectively reduce the complexity of data, which has attracted extensive attention [6]. Its main technical principle is to decompose complex data into several relatively single modal components according to different spectra. Subsequently, its derivative methods and other similar treatment methods are also studied and developed, such as ensemble empirical mode decomposition (EEMD) [7], complementary ensemble empirical mode decomposition(CEEMD) [8], variational mode decomposition (VMD) [9], wavelet transform (WTB) [10], seasonal and trend decomposition using loess (STL) [11], etc. To be sure, by decomposing the complex data into several components with their own frequencies, the complexity of the data is greatly reduced, thereby improving the prediction accuracy significantly. Regrettably, the randomness hidden in the residual and high-frequency terms cannot be well mined and processed, which leaves the trouble and thorny obstacle to accurate prediction.

In the existing research, many scholars have paid attention to the analysis of residual terms, including quadratic decomposition strategy [12], error factor correction technology [13] and so on. However, excessive decomposition of high frequency terms may change the original characteristics of randomness. As far as we know, few studies investigate the remainder after the exclusion of structural components (trend and seasonality variations) from the perspective of residual component disposing. Whereas, in this study, it is believed that some predictive information remains in the remainder and if discarding them thoroughly would certainly exert a negative influence on prediction accuracy. The technical advantage of this idea is that it can truly describe the original appearance of randomness. The HSR passenger flow data involving individual travel behavior shows obvious randomness. When we separate the structural components from the original data, including the trend term and the period term, the remaining residual term covers the main random factors. It is easy to learn the evolution law hidden in the structural components except the residual term. How to effectively utilize the random factors contained in the residual is the key problem to obtain accurate prediction. To a great extent, we believe that the weak but indispensable information is still contained in the remainder component, and some random incidents occurred in the past may also happen in the future. Based on such a theoretical consideration, moving block bootstrap (MBB) is employed to bootstrap the residual series and generate a large number of similar samples, so as to simulate the stochastic feature of future series and improve the reliability of the data.

As a way to fill the gap in this issue, this paper develops a new hybrid system with deterministic and probabilistic forecasting capacities for HSR passenger demand. The data is firstly decomposed into trend, seasonality components, and residual by the seasonal and trend decomposition using loess (STL). Then, the residual is bootstrapped by moving block bootstrap (MBB), and the obtained information fragments is disposed to construct one thousand similar samples by virtue of extrapolating the disaggregated subseries, thereby simulating the stochastic feature of future series. After the extrapolation procedure, the generated similar deseasonalised (S) and detrended (T) subseries are predicted by seasonal autoregressive integrated moving average (SARIMA), kernel extreme learning machine (KELM), respectively. And support vector regression (SVR) is used in similar residuals prediction, including detrended series (TR) and deseasonalised series (SR). Similar global forecasts are obtained according to the reaggregation of the extrapolations. Finally, bagging predictors is performed to aggregate the obtained similar forecasts to get the final forecast results. To provide more profound advice for forethoughtful planners of railway operation department, we give deterministic and probabilistic forecasts of HSR passenger demand from one-step to three-step.

The rest of this paper is composed as follows. Sections Section 2 gives the literature review revolving around passenger demand of HSR forecasting. Sections Section 3 introduces the architecture of our established system and elucidates the detailed information of models and strategies adopted in this paper. In the  4th Section, the experiment setup and forecasts analysis are presented in detail. Finally, Section 5 represents the summary of this paper and further work.