Process-Aware Procurement Lead Time Prediction for Shipyard Delay Mitigation

Accurately predicting procurement lead time (PLT) remains a challenge in engineered-to-order industries such as shipbuilding and plant construction, where delays in a single key component can disrupt project timelines. In shipyards, pipe spools are critical components; installed deep within hull blocks soon after steel erection, any delay in their procurement can halt all downstream tasks. Recognizing their importance, existing studies predict PLT using the static physical attributes of pipe spools. However, procurement is inherently a dynamic, multi-stakeholder business process involving a continuous sequence of internal and external events at the shipyard, factors often overlooked in traditional approaches. To address this issue, this paper proposes a novel framework that combines event logs, dataset records of the procurement events, with static attributes to predict PLT. The temporal attributes of each event are extracted to reflect the continuity and temporal context of the process. Subsequently, a deep sequential neural network combined with a multi-layered perceptron is employed to integrate these static and dynamic features, enabling the model to capture both structural and contextual information in procurement. Comparative experiments are conducted using real-world pipe spool procurement data from a globally renowned South Korean shipbuilding corporation. Three tasks are evaluated, which are production, post-processing, and procurement lead time prediction. The results show a 22.6% to 50.4% improvement in prediction performance in terms of mean absolute error over the best-performing existing approaches across the three tasks. These findings indicate the value of considering procurement process information for more accurate PLT prediction.

💡 Research Summary

The paper tackles the persistent problem of forecasting procurement lead time (PLT) for pipe spools in shipbuilding, an engineered‑to‑order sector where a single delayed component can cascade into costly schedule overruns. Existing approaches rely exclusively on static spool attributes such as diameter, weight, or material, ignoring the fact that procurement is a dynamic, multi‑stakeholder process that generates a rich stream of time‑stamped events. To bridge this gap, the authors propose a process‑aware prediction framework that fuses static spool data with features extracted from the procurement event log.

The event log records every activity in the procurement workflow—material release, work order issuance, fabrication steps, inspections, hand‑overs, transportation, and receipt—each with a case identifier, activity name, timestamp, and auxiliary attributes. From each event the authors derive three temporal descriptors: elapsed time since the start of the case, lagged time since the previous event, and the day‑of‑week. These descriptors capture both the continuity of the process and its weekly cyclicity.

The modeling architecture treats the two data sources separately before merging them. Dynamic, time‑related features are fed into a bidirectional Long Short‑Term Memory (Bi‑LSTM) network, which learns contextual representations from both past and future events within a trace. Static spool attributes are passed through a multi‑layer perceptron (MLP) with ReLU activations, producing a dense vector that encodes the physical characteristics of the component. The outputs of the Bi‑LSTM and the MLP are concatenated and processed by a final fully‑connected layer that regresses the PLT. The loss function is mean absolute error (MAE), and early stopping based on a validation set prevents over‑fitting.

The experimental evaluation uses a real‑world dataset from a leading South Korean shipyard: 106,403 pipe spools, 12 static attributes, 5 dynamic attributes, and a total of 2,919,961 logged events, with trace lengths ranging from 18 to 36 steps. The data split follows a 70/10/20 train/validation/test scheme. Baselines include traditional static‑only models (Decision Tree, Random Forest, Linear Regression, MLP) and a suite of sequential deep models (RNN, LSTM, GRU, Bi‑GRU, Transformer). Performance is measured by MAE, RMSE, MAPE, and computational cost (inference time).

Results show that the proposed hybrid model consistently outperforms all baselines across three prediction tasks: production lead time, post‑processing lead time, and overall procurement lead time. For PLT, the model achieves an MAE of 4.66 days and RMSE of 7.13 days, representing a 50.4 % reduction in MAE and a 45.7 % reduction in RMSE compared with the best static baseline (Decision Tree). Similar gains (22.6 %–38.5 % MAE reduction) are observed for the other two tasks. Among sequential architectures, Bi‑LSTM delivers the best balance of accuracy and efficiency, while Transformers, despite their popularity, incur higher computational cost with modest accuracy.

Ablation studies confirm the critical contribution of both feature groups. Removing temporal descriptors (TRF) degrades MAE by roughly 30 %–45 % across tasks; removing the entire event‑log information (EL) leads to even larger drops, with PLT MAE increasing from 4.44 to 10.70 days. These findings underscore that the temporal dynamics and contextual cues embedded in the event log are indispensable for reliable PLT estimation.

Computationally, the hybrid model requires more processing time than static‑only methods (approximately 458 s for production, 1025 s for post‑processing, and 794 s for PLT inference on the test set), but the authors argue that the accuracy improvements justify the overhead in an industrial setting where schedule certainty is paramount.

In conclusion, the study demonstrates that incorporating process‑aware information—specifically, event‑log derived temporal features—into a deep learning framework yields substantial predictive gains over traditional static‑only approaches. The work paves the way for more sophisticated, real‑time decision support tools in shipbuilding and other engineered‑to‑order domains, and suggests future extensions such as multi‑project interaction modeling, external supply‑chain volatility integration, and deployment of online learning pipelines.