Conference Proceedings on the Web 22: Transforming the Marine Transportation System Through Multimodal Freight Analytics (2018)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

51 Student Honor Session Student Honor Panel Presentations Helen Brohl, U.S. Committee on the Marine Transportation System, presiding Assessing the Impact of Collaboration of Truck Carriers in Intermodal Freight Transportation Majbah Uddin, University of South Carolina, presenter n recent years, intermodal transport has become an increasingly attractive alternative to freight shippers, and this trend is likely to continue as state and federal agencies implement policies to induce a freight modal shift from road to intermodal transport. However, the current intermodal freight transport is not as efficient as it could be. According to the National Private Truck Council, about 28 percent of truck trips are empty. In intermodal transport, an empty container often needs to be transported from the empty container depot to the shipper, and conversely, an empty container needs to be transported from the consignee to the empty container depot. These empty truck movements decrease the motor carrier’s profit, increase traffic congestion, decrease roadway safety, and add unnecessary emissions to the environment. This study investigated a potential collaboration strategy for motor carriers to reduce the number of empty trips to benefit the participating members economically and the general public through reduced impacts. In addition, this study developed a binary integer linear program to determine the optimal collaboration plan that would maximize the carriers’ profit. A practical constraint considered by the model is that a carrier would collaborate only if its profit obtained under collaboration were higher than or equal to its profit without collaboration. Therefore, this study considered several operational characteristics, such as number of trucks owned by a carrier, time window constraints at the intermodal terminal, and shipper/consignee. Furthermore, this study considered uncertainty in time windows. The proposed model/collaboration strategy was tested on an actual intermodal network and was demonstrated to be effective in meeting all shipment constraints while maximizing the profit of participating members. I

52 Determining the Accuracy of Vessels’ Estimated Time of Arrival with Different Output Parameters Akintola Aremu, Texas Southern University, presenter ultimodal transportation is a growing field and a strategy in the global supply chain. Therefore, it is important to look at ways and technologies that can be employed to improve the effectiveness of this system. One of the major challenges faced in multimodal transportation is the variability in the time of arrival of the different modes of transportation involved. Although trucking companies may have access to some levels of information to track the vessels and cargos, the companies do not have accurate estimations for the arrivals of those vessels to the port. Providing an accurate time of arrival will improve the planning and scheduling of processes and thereby reduce down time, delay, and cost. This study applied machine learning methods to estimate arrival times of vessels and compared results obtained from two output parameters, actual time of arrival and delay, which can be used to determine the estimated time of arrival afterwards. One major factor that determines the effectiveness of a multimodal system is the level of synchronization of the time of arrival of participating modes of transportation. A synchronized system can be achieved through effective communication, the use of tracking technologies, and analysis of historical data. Besides the use of statistical analysis such as regression and the study of historical patterns, machine learning has proved to be an effective way of predicting the estimated time of arrival of vessels. Research has predicted the estimated time of arrival of vessels through the use of supervised machine learning, including decision trees and support vector machine artificial neural networks. In the use of supervised machine learning algorithms, techniques—including algorithm tuning, spot check algorithm, data filtration, and normalization—are employed to improve the performance of the algorithm and to derive accurate results. This work evaluated the impact of an output parameter change on the accuracy of prediction obtained. The relevant data included latitude, longitude, distance to destination, current speed of the vessel, change in speed over the past 3 hours, average speed based on the past 12 hours, time used for calculating the average speed, length of the ship, breadth of the ship, and the estimated time of arrival (ETA) of the ship’s agent (acquired from AIS). Other quantifiable factors specific to the port of concern are considered and incorporated into the system to improve the accuracy of the result. The output parameters for this work are actual time of arrival (ATA) and the delay as calculated by Equations 1 and 2, respectively. ATA = ((DATA – Dcollection) ∗ 1,440) + ((HATA – Hcollection) ∗ 60) + (MATA – Mcollection) [1] delay = ((DATA – DETA) ∗ 1,440) + ((HATA – HETA) ∗ 60) + (MATA – META) [2] where M

53 Dcollection = data collection day (mm/dd/yyyy), DATA = arrival day (mm/dd/yyyy), Hcollection = data collection hour (hours), HATA = arrival hour (hours), Mcollection = data collection minute (minutes), MATA = arrival minute (minutes), DETA = estimated day of arrival (mm/dd/yyyy), HETA = estimated hour of arrival (hours), and META = estimated minute of arrival (minutes). During processing, records with missing or incomplete data and outliers were eliminated. The data were normalized, and important parameters were identified by applying Garson’s algorithm. With the use of a neural network with 15 input neurons, one hidden layer with eight neurons and one output layer (ETA), ETA was predicted. Because of the uniqueness of the study area (Port of Houston), ETA was determined through a combination of ETA to the pilot point and ETA to the terminal. The process was repeated with delay as output. The results were compared by finding the mean absolute error (MAE) and root mean square error (RMSE) as shown in Table 2. TABLE 2. MAE and RMSE from ETA Predictions Using Actual Time of Arrival and Delay as Output Parameters. Parameter and Variable Time To Pilot Point To Buoy Total Error ATA MAE 0.086111 0.176388 0.262498 RMSE 0.148524 0.208813 0.357337 Delay MAE 0.062686 0.356215 0.418901 RMSE 0.062686 0.460897 0.523583 The result shows that using delay as an output variable performed better than ATA when considering the trip from last port of call to the pilot point. However, ATA outperformed the use of delay as output parameter for trips from the pilot point to the terminal. This poor performance is likely because of inconsistencies and unpredictability that may arise from the culture of the port, such as daylight restrictions, dredging, pilot policies, and restrictions due to vessel size.

54 Freight Mode Choice Modeling Using the 2012 Commodity Flow Survey and Longitudinal Business Data 2012 Lokesh Kumar Kalahasthi, Rensselaer Polytechnic Institute, presenter José Holguín-Veras, Shama Campbell, Diana Ramírez-Ríos, and Jeffrey Wojtowicz, Rensselaer Polytechnic Institute Carlos A. González-Calderón, National University of Colombia at Medellin his session summarized the key findings from the National Cooperative Freight Research Program (NCFRP) Project 44, Impacts of Policy-Induced Freight Modal Shifts. This research used the 2012 Commodity Flow Survey (CFS) confidential microdata to estimate freight discrete and continuous mode choice models at the national level with more than 2 million observations. The objective of this project was to analyze the factors influencing freight mode choice and to develop analytical methods to quantify the impacts of various policies on modal patterns. It is important for public policies to encourage effective use modes of transportation that minimize externalities while fostering economic growth. To achieve this goal, it is crucial to have a better understanding of the variables and behavior of freight agents that influence mode choice decisions. However, there is a lack of research in the field of freight mode choice. This study helps to fill this gap by providing insights into freight mode choice, enabling more accurate demand forecasts, better quantification of modal split and respective impacts on society. The research provides policymakers with tools to assess the effects of policies on freight mode share. DATA PREPARATION A major challenge is the lack of a single data source for estimating freight mode choice models. The CFS data comprise shipment attributes (weight, value, mode, commodity type, origin, destination), but do not have modal attributes (travel times, costs, distances). CFS has distance by only one mode of transport, the mode selected. This factor does not help in modeling, as distance by both modes (rail and truck) is required. CFS data are combined with the Longitudinal Business Database (LBD) for establishment attributes, while HERE (https://www.here.com/en) data from FHWA are used to estimate truck distances and travel times; 2012 confidential waybill data from the Federal Railroad Administration (FRA) are used to obtain rail rates and travel times; and rail network data from FRA are used to obtain rail distances between zip codes in the United States. Because of data limitations, this study focused on estimating freight mode choice only between rail and truck. Truck costs were estimated as a function of shipment size, distance, and travel time from HERE data with the use of the cost models available in the literature. Rail rates and travel times were inferred from the waybill data and FRA rail network data. In summary, the preparation of the final data set required the use of the following data sets/models from these data sets: CFS, LBD, HERE, waybill, FRA rail network, and truck cost models. T

55 METHODOLOGY The research involved both qualitative and quantitative techniques. As part of the qualitative effort, the team conducted in-depth interviews with freight agents. The companies represented in these interviews comprised five shippers, four receivers, and two carriers. The participant firms were of various sizes that operated more than 5,500 establishments and employed more than 1.3 million workers in the United States. As part of the quantitative effort, the team estimated aggregate and disaggregate models using the CFS data. Aggregate models were ordinary least square regressions between probabilities and the average modal attributes. The disaggregate models were discrete/continuous models with shipment size as endogenous variable. The greater circle distance was used as an instrument to correct for endogeneity between mode and shipment size. Two types of weighted models that considered the mode share for domestic cargo only and total cargo, including imports and exports, were estimated. The weights adjusted the sample to represent the actual market mode share provided by the FAF. KEY FINDINGS FROM THE IN-DEPTH INTERVIEWS The interviews obtained information on supply chain operations, factors influencing freight mode choice, and suggestions to improve sustainability. Major factors influencing mode were consistency in delivery time, reliability, customer care, cost, and commodity. A small-scale shipper mentioned that carriers of comparable size to their company provide better services because the shipper’s business represented a sizable portion of the carrier’s revenues. Suggestions to promote sustainability were to increase truck weight limit over 80,000 pounds, provide efficient switching between railroads, increase manifest rail operations, provide better geographic spread of intermodal services, provide storage/refrigeration/packaging services in rails, dredge canals, maintain locks and dams, and preserve land along rivers and waterways. SELECTED FREIGHT MODE CHOICE MODELS Using CFS data, the team estimated more than 1,000 econometric models— 266 aggregate models and 762 disaggregate models—that predicted the choice of mode by commodity type at two-digit Standard Classification of Transported Goods (SCTG) codes. Two types of variable combination—cost and travel times and generalized costs—took into account the value of commodity and the opportunity cost associated with travel time. Table 3 shows the discrete choice models with cost (C) and travel time (TT) as variables. The utilities of truck and rail are Ut = β0 + βc Ct + βTtTTt + εt and Ur = βcCr + βTtTTr + εr, respectively. On the basis of the value of time (VOT) in dollars per hour for an average shipment, paper has more VOT as compared with transportation equipment and pharmaceutical products.

56 TABLE 3. Disaggregate, Cost and Travel Time. SCTG (2-digit) Description β0 z-Stat. βc z-Stat. βTt z-Stat. Obs. Wald(2) p-Value ρ VOT 37 Other trans- portation equipment 3.18 43.52 –2.67E-03 –14.94 –3.52E-03 –13.50 22,000 282.70 0.00 0.10 0.76 21 Pharmaceutical products 5.82 47.64 –1.12E-03 –2.24 –9.70E-04 –1.68 60,000 48.27 0.00 0.02 1.15 28 Paperboard articles 5.09 90.98 –9.94E-04 -–3.72 –3.20E-04 –4.58 120,000 1670.00 0.00 0.16 3.11 Table 4 shows the discrete choice models of the top five sensitive commodities to generalized cost at 5 percent opportunity cost. An opportunity cost of 5 percent was found to be significant compared with 10 percent and 25 percent. All commodities have positive constants, showing more inclination toward trucks. TABLE 4. Disaggregate Domestic Weights, Generalized Cost 5%. SCTG (2-digit) Description β0 z-Stat. βg5 z-Stat. Obs. Wald(<15) p-Value ρ 28 Paperboard articles 3.42 63.99 –4.66E-04 –10.65 120,000 113.40 0.00 0.05 17 Gasoline, aviation fuel, and ethanol 3.96 48.05 –3.57E-04 –12.80 40,000 163.80 0.00 0.45 39 Furniture and lighting equip. 6.97 69.53 –3.26E-04 –6.17 110,000 38.02 0.00 0.05 11 Natural sands 3.57 115.80 –1.52E-04 –12.63 34,000 159.60 0.00 0.14 26 Wood products 3.53 146.30 –1.42E-04 –51.64 210,000 2666.00 0.00 0.13 CONCLUSION These freight mode choice models are the first to use CFS microdata with 2 million observations. The team is finalizing the models and applying them to real-world case studies. A guidebook for public practitioners was prepared to assist with the use of these models in estimating mode shifts due to different policies. The guidebook also provides a detailed explanation of factors influencing mode choice, a review of other modeling techniques, such as inventory theory and supply chain management, the best-performing models from the CFS, and examples of how to apply these models using case studies. These freight mode choice models along with freight generation models serve as a potential tool for nationwide freight planning.