PaperlayerView exact revision on arXiv

Source-linked AI summary

Travel Time Prediction in Supply Chain Management Using Machine Learning

Balaji Venkateswaran

arXiv:2609.38190v1cs.LGcs.AI

TL;DR

The paper addresses travel-time prediction for supply-chain shipments, where arrival estimates matter for operations and customer satisfaction. It applies a 1D-CNN to vessel transit data and reports a 7% mean absolute percentage error.

  • Problem

    Transport tracking systems report planned movement and current location, but not how subsequent transport is likely to unfold; delays can cascade and cause missed connections.

    Source: §2.6 · p. 11

  • Method

    The study applies a 1D-CNN regression model to vessel transit data to predict travel time.

    Source: §6.4 · p. 37

  • Results

    7% mean absolute percentage error was obtained for vessel travel-time prediction.

    Source: §6.4 · p. 37

  • Takeaways & Limitations

    A wrapper application can supply route features for point-to-point travel-time prediction, while deciding the path separately.

    Source: §5.1 · p. 35

    The authors describe the model as basic and suggest adding and studying features and collecting three years of data.

    Source: §5.2 · p. 36, §6.3 · p. 37

Abstract

from arXiv · show

The purpose of this research is to find data and methods using machine learning and deep learning to correctly predict the estimated travel time for transportation and logistics in a supply chain system. The supply chain ecosystem is very complex and heavily relies on the transportation and logistics of raw materials and finished goods. Accurate travel time estimation is critical because it helps supply chain members to improve logistics consistency and performance. This helps in planning, demand forecasting, lead time management and assembly planning. The logistics on the delivery side of the customer also plays a crucial role in customer satisfaction and voice of customer. With the collection of huge historical data and using novel techniques, the research builds an accurate model to predict travel time of inventory.

1 Introduction

Accurate freight travel-time estimates support supply-chain planning and logistics performance. Forecasting is difficult amid multimodal complexity and variable travel conditions.

1.2 Research Problem

  • Arrival timing shapes manufacturing and assembly planning and downstream customer satisfaction, while prior ETA studies often used linear regression despite complex datasets.The study frames deep-learning methods as an alternative for complex ETA prediction.Source: §1.2 · pp. 4–5

1.3 Research Aims

  • The research aims to apply advanced deep-learning techniques to predict travel time for transportation and logistics in supply chains.Source: §1.3 · p. 5

1.1 Research Background and Scope

  • The study targets waterway cargo-ship shipments, against a background of logistics complexity and multiple factors affecting travel time.The text names temperature, traffic, driver, direction, and transportation relationships as relevant factors.Source: §1.1 · p. 4, §1.3 · p. 5

1.4 Research Objectives

  • The objectives cover online shipment, weather, and geolocation data; cleaning and aggregation; feature selection; and 1D-CNN feature-map generation.Feature selection may use principal component analysis or a random forest regressor.Source: §1.4 · p. 5

1.5 Purpose of Research

  • The purpose is to develop an accurate ETA model from historical data for estimating future shipment travel times.Source: §1.5 · p. 5

1.6 Significance of the Study

  • The study links travel-time prediction to supply-chain interruptions involving extreme weather and traffic congestion, and to modeling transit journey-time features.Source: §1.6 · p. 5

1.7 Research Design

  • The research focuses on machine-learning travel-time estimation amid limited work on freight CNN or LSTM prediction and long-term multimodal forecasting.Freight monitoring, for example with mobile sensors attached to transported goods, is also described as necessary.Source: §1.7 · p. 6

1.8 Structure of the Paper

  • The paper reviews prior work, presents its methodology and results, and then discusses limitations, future work, implications, and recommendations.Source: §1.8 · p. 6

2 Literature Review

The literature review examines accurate travel-time and ETA prediction across freight logistics, navigation, and ride-hailing, including its role in logistics planning and transportation decisions and the use of machine learning to estimate travel time by route and departure time.

2.2 Research Objectives in Related Work

  • Existing ETA studies report varied machine-learning results, including Random Forest’s usefulness but complexity, DNN gains, SVR’s advantage over two alternatives, and shipping-model accuracy above baseline.Yuan et al. (2020) report a 4.82% improvement; Servos et al. (2019) report errors up to 17 hours over travel periods of up to 30 days and a need for iterative retraining. Viellechner and Spinler (2020) report over 77% accuracy versus 59% for the baseline. XGBoost has the least MAPE at 11.96%, 9.30% lower than other models generally.Source: §2.2 · pp. 7–8

2.3 Travel Time Prediction

  • Travel-time prediction is difficult because traffic, weather, vehicles, routes, and transport relationships affect journeys, while short-term methods use traffic-state data or statistical approaches.Accurate prediction also supports advanced traffic information systems, traffic management, and logistics planning.Source: §2.3 · pp. 8–9

2.4 Machine Learning in Travel Time Prediction

  • The review argues that machine learning suits freight’s complex, nonlinear behavior, and reports SVR as the most accurate method in its use case, outperforming average-based techniques.A DBN approach stacks Restricted Boltzmann Machines to learn traffic features unsupervised, then uses sigmoid regression for supervised travel-time prediction; it can train across an entire road network.Source: §2.4 · pp. 9–10

2.5 The Travel Time Prediction Problem

  • Travel-time prediction must contend with congestion, sparse comparable journeys, and computational growth as datasets and model complexity increase.Similarity choices affect sparsity, while algorithm estimates may need simplification to improve speed and scalability (Wu et al., 2020).Source: §2.5 · pp. 10–11

2.6 Estimated Time of Arrival (ETA) in Supply Chain Management

  • Intermodal ETAs matter because an early-leg delay can cascade into missed connections, while timely estimates at transshipment points can enable action to compensate for delays.Intermodal chains combine scheduled and unscheduled transport modes, and reliable ETA estimation requires large, rapidly collected datasets.Source: §2.6 · p. 11
  • For intermodal ETA prediction, differences among modes and parties complicate integrated estimates, and multimodal schedules and capacity constraints can create peaks in arrival-time distributions.Data gathering and integration are harder because different parties use different aims and IT systems; transshipment capacity also constrains predictions.Source: §2.6 · p. 11

2.7 Machine Learning Approach Estimation of Travel Time

  • A machine-learning ETA model for intermodal chains should represent logistics structure and split prediction into individual transport-leg subproblems before combining estimates across the chain.Without that structure, transport, buffer, and storage times can be blended, making ETA estimates meaningless; legwise predictions can be passed forward and assembled into an overall forecast.Source: §2.7 · p. 12

2.9 Research Gap

  • The review identifies data availability and quality, end-to-end logistics context, and downstream delay impacts as priorities for useful machine-learning ETA support.ETAs can support operational decisions about people, vehicles, tools, and infrastructure, while logistic-node and process information matters even when the full process is not modelled in detail.Source: §2.9 · p. 14

2.10 Conclusion

  • ETA informs navigation and ride-hailing decisions, including route discovery and vehicle dispatch, while reliable predictions can reduce travel costs, electricity demand, and vehicle emissions (Jian Li, 2020).Source: §2.10 · p. 14
  • A novel machine learning algorithm can estimate travel time for a specific route and departure time (Wang et al., 2018).Source: §2.10 · p. 14

3 Methodology

The methodology combines shipment data collection and preprocessing with feature reduction and a deep-learning pipeline for predicting estimated arrival times. Its described workflow uses 1D-CNN feature maps followed by LSTM processing, with convolution, pooling, and fully connected layers forming the CNN architecture.

3.1 Methodology

  • The proposed pipeline passes selected features through a 1D-CNN to create feature maps, then uses LSTM processing to produce the shipment’s projected arrival time.The methodology also describes exploratory data analysis and feature selection before modeling.Source: §3.1 · pp. 14–15

3.2 Outline of This Section

  • The methodology outlines data collection, feature selection, PCA, deep-learning prediction, and 1D-CNN as its main stages.Its section outline identifies these topics as the sequence of methodological components.Source: §3.2 · p. 15

3.3 Data Collection for Estimated Time Arrival in Supply Chain Management

  • The study collects shipment-delivery data covering geolocation, time, weather, and shipment details, using Eesea and AIS sources for shipment information.Eesea data is described as paid maritime intelligence data, while AIS data is non-public and includes shipment IDs and distances travelled.Source: §3.3 · pp. 15–16

3.4 Data Pre-processing

  • Preprocessing addresses incomplete, duplicated, and outlier records by removing features with many nulls, imputing remaining missing values, and removing duplicates and selected outliers.Mean and median values are used to fill features with fewer missing elements.Source: §3.4 · p. 16

3.7 Principal Component Analysis

  • Dimensionality reduction removes variables to simplify learning and reduce computation time, while PCA is used to retain features accounting for more than 95% of data variance.The number of components can be adjusted as a modeling hyperparameter to compare prediction accuracy and error.Source: §3.6 · p. 16, §3.7 · p. 17

3.9.1 Convolutional Layer

  • The 1D-CNN uses cascaded convolutional layers whose outputs are activated by ReLU, with extracted features becoming more abstract across convolution kernels.The study describes two convolutional layers and uses the output of each as input to the next.Source: §3.9.1 · p. 19, §3.9.4 · p. 20

3.9.2 Pooling Layer

  • Maximum pooling discards some features, and the fully connected layer produces the final prediction using a nonlinear function while backpropagation updates network weights to reduce error.The fully connected layer aggregates neurons into a one-dimensional vector for faster processing.Source: §3.9.2 · p. 19, §3.9.3 · p. 19

3.9.4 Activation Function

  • The activation function applies nonlinear mappings to extracted features; the study names ReLU, Sigmoid, and Tanh and describes a CNN-to-regression pipeline for ETA prediction.The model configuration specifies two hidden convolutional layers using ReLU; selected data is transformed into feature maps before regression.Source: §3.9.4 · pp. 19–20

4 Results

The results describe the transit dataset, feature analysis and preparation, and a 1D-CNN model reported at 7.10% MAPE on data from more than 13,000 vessels.

4.1 Exploratory Data Analysis (EDA)

  • The 20-million-plus-record, 18-feature dataframe is numeric and has no nulls; TIME_TAKEN clusters around 1,500–2,000 seconds and is close to normally distributed.LAT_NEXT, SPEED, HAVERSINE_DISTANCE, and TIME_TAKEN are described as highly variable, motivating the authors’ plan to use PCA to address uneven distributions.Source: §4.1 · pp. 20–21

4.2 Correlation Study

  • For HAVERSINE_DISTANCE ≤ 2,500, TIME_TAKEN and distance are highly correlated, with the relationship varying alongside vessel speed and snapshot conditions.The scatter plot presents this relationship.Source: §4.1 · p. 21, Figure 13 · p. 26

4.3 Additional Features

  • The derived features include next-position latitude and longitude, plus HAVERSINE_DISTANCE, the spherical distance in kilometers between two geographic points.These features are introduced to extract more information from the available data.Source: §4.3 · p. 26

4.4 Simple Linear Regression

  • Simple linear regression feature importance led the authors to drop NAME_ENCODED, DESTINATION_ENCODED, and IMO as having no impact on TIME_TAKEN.Figure 15 presents the feature-importance values.Source: §4.4 · p. 28, Figure 15 · p. 29

4.5 Principal Component Analysis

  • PCA produced 14 components; PC13 and PC14 were dropped for having no explained variance, leaving 12 components for the 1D-CNN regression.The explained-variance distribution is shown in Figure 17.Source: §4.5 · p. 28, Figure 17 · p. 30

4.6 1D-CNN Deep Learning Model

  • The 1D-CNN uses two hidden ReLU convolutional layers on scaled data; training on 80% of the dataset for 10 epochs ended with validation loss 0.3329.Training used batch size 1,000; the authors report that validation loss decreased considerably across epochs.Source: §4.6 · p. 31

4.7 Conclusion

  • 7.10% MAPE: the 1D-CNN is reported as accurate on a year of transit data from 13,000+ vessels and 22+ million datapoints.The authors say the model can support a wrapper application for point-to-point travel-time predictions.Source: Table 4 · p. 31, §4.7 · p. 31

5 Discussion

The model predicts segment travel time from supplied route features, while the calling application chooses the path and sums segment times for a long journey. The discussion proposes vessel load, nearby traffic, and weather as possible additions.

5.1 Discussion of Results

  • 7% MAPE was obtained using the features listed in Table 5.The TIME_TAKEN target is measured in seconds.Source: §5.1 · p. 35, Table 5 · p. 35
  • The calling application chooses the route, and the model predicts travel time between points A and B from the supplied features.For a long journey, the route is split into parts and the predicted segment times are summed.Source: §5.1 · p. 35, Table 6 · p. 35

5.2 Improvements Suggested

  • The proposed additional inputs are vessel load, ships within 50 km at the snapshot location, and weather at that location and time.Actual weather could be used for training and predicted weather for future inputs.Source: §5.2 · p. 36
  • The authors characterize possible gains from these added features as marginal; the existing model is described as predicting within ±7% of actual time for long-haul vessel journeys.The passage does not report separate results for models with the proposed additions.Source: §5.2 · p. 36

6 Summary, Implications, and Recommendations

The study builds a 1D-CNN regression model for vessel travel time from collected and processed vessel data. It reports 7% MAPE and recommends further data and feature work and possible model combinations.

6.1 Summary

  • The study collected 23 million 2023 datapoints for transits of 13,000+ vessels, cleaned the data, engineered and selected features, and used PCA to retain 12 features for the 1D-CNN.The summary identifies 15 API features initially, then 17 after feature additions, 14 after removing three features, and 12 after PCA.Source: §6.1 · p. 36

6.2 Implications

  • The paper presents 1D-CNN as a way to identify data-feature correlations that may be difficult to detect with other methods, with potential applications in shipping-route optimization and logistics planning.Source: §6.2 · p. 37

6.3 Recommendations for Future Research

  • Future research recommendations include collecting more data covering three years and adding and correlating new features, including weather, vessel traffic, and tonnage.Source: §6.3 · p. 37
  • The authors propose combining the 1D-CNN with LSTM or attention networks to further improve travel-time prediction accuracy.Source: §6.3 · p. 37

6.4 Conclusion

  • The vessel travel-time prediction achieved a mean absolute percentage error of 7%, and the 1D-CNN converged faster with the large dataset used.The model was implemented for the regression problem.Source: §6.4 · p. 37

A.1 Getting List of Vessels by IMO

The appendix describes vessel-list retrieval and the collection of cargo-vessel tracking data for 2023. History data is requested in intervals because each API hit can return only 30 days.

A.1 Getting List of Vessels by IMO

  • The vessel-list API returns a zip file of tracked vessels, with its CSV columns described in Table 7.The API is subscription-based and costs EUR 649 per month.Source: §A.1 · p. 37, Table 7 · p. 38

A.2 Getting Travel Geo-coordinates for One Year

  • The authors filtered a sample of 13,000 IMOs for cargo vessels and gathered their 2023 geo-positional activity through API requests.Source: §A.2 · p. 39
  • Because each history API hit retrieves at most 30 days, the authors made 12 requests per vessel IMO to construct the tracking dataframe.The requested data includes positions and vessel identifiers and metadata.Source: §A.2 · pp. 39–40

B.1 Feature Engineering

Feature engineering filters and reshapes vessel-position records into a travel-time modeling dataset. It derives next-position coordinates, timestamp features, Haversine distance, and elapsed time, then applies encoding and cleaning steps.

B.1 Feature Engineering

  • Vessel records with speed ≤1 knot are excluded, and SHIP_TYPE plus several unique identifiers are dropped while IMO is retained.The retained identifiers include IMO; MMSI, UUID, and ENI are removed.Source: §B.1 · p. 40
  • Latitude and longitude lead values are grouped by IMO and DESTINATION to represent the next position, and timestamp fields are expanded into year, month, day, hour, and minute.The section also describes calculating Haversine distance between GPS points.Source: §B.1 · pp. 40, 42
  • TIME_TAKEN, measured in seconds, is the prediction target, and only values ≤3000 are retained to avoid outliers..Source: §B.1 · p. 42
  • Categorical values are label-encoded, while missing COURSE and HEADING values are set to zero and rows missing TIME_TAKEN are removed.The encoded columns are COUNTRY_ISO, NAME, TYPE_SPECIFIC, and DESTINATION.Source: §B.1 · p. 42
  • Haversine distance is computed between GPS positions using latitude and longitude in radians and an Earth radius of 6371 kilometers.The resulting distance is returned in kilometers.Source: §B.1 · pp. 42–43

C.1 Simple Linear Regression

The section fits a simple linear regression model to predict travel time and evaluates its predictions on a held-out test set.

C.1 Simple Linear Regression

  • A linear regression model is trained on a 70/30 train-test split with random_state=42 and evaluated using mean squared error and MAPE.The supplied passages do not report the resulting metric values.Source: §C.1 · p. 43

C.2 Principal Component Analysis

The PCA workflow standardizes the data, transforms the features into principal components, and selects nine components for the final transformation.

C.2 Principal Component Analysis

  • The data are standardized before PCA, and the transformed features are reduced to nine principal components.The code also computes explained-variance ratios and plots cumulative explained variance against the number of components.Source: §C.2 · pp. 43–44

C.3 1D-CNN Deep Learning

The paper trains a 1D-CNN on PCA-transformed inputs and evaluates its predictions using RMSE-related and MAPE calculations. The supplied passages describe the model and evaluation procedure but do not provide metric outcomes.

C.3 1D-CNN Deep Learning

  • PCA-transformed data are divided into training and testing sets; the model is trained on the training data for 10 epochs with batch size 64, using the test set for validation.Source: §C.3 · p. 44
  • The 1D-CNN uses Conv1D layers with 64 and 128 filters, max pooling, dense layers of 128 and 64 units, and a single linear output.The first convolution uses a kernel size of 3 and ReLU activation; the model is compiled with Adam and mean squared error loss.Source: §C.3 · p. 44

C.4 RMSE and MAPE

  • The CNN is evaluated on the test set, and predictions and targets are inverse-transformed before mean squared error and MAPE are calculated.MAPE excludes observations whose true value is zero.Source: §C.3 · p. 44, §C.4 · p. 45
Loading 2609.38190v1…