University of Twente Student Theses


Hierarchical forecasting of engineering demand at KLM Engineering & Maintenance

Breed, I (2019) Hierarchical forecasting of engineering demand at KLM Engineering & Maintenance.

[img] PDF
Abstract:Engineering demand for KLM Engineering and Maintenance consists of a varied collection of tasks for different customers and aircraft on routine and non-routine basis. Every combination of these characteristics presents a unique demand subset that in turn has its own behaviour. A total of 1716 different subsets of demand are present each of which could require its own forecast. Accurately forecasting these different series has proven to be challenging with the current forecasting approach. Leading to difficulties in accuratly matching engineering capacity to demand We propose a framework that applies multiple models and leverages the benefits of forecast combination and reconciliation. Different forecasting models are able to extract different information from the demand subset resulting in different forecasts. Using multiple models ensures that we are able to extract sufficient information from all the different subsets without risking model miss-specification. Combining the resulting forecasts creates forecasts that combine the information extracted by the different models. By leveraging multiple sources of information combined forecasts are more accurate than single forecasts. We then select the most accurate forecast through the mean absolute scaled error (MASE), resulting in a single best forecast for each of the demand subsets. Each of the 1716 different subsets end up with an individual forecast but these are incoherent between the levels of aggregation. The forecasts for each aircraft type demand do not sum equal to the forecast of total demand. Both the sum and total demand represent the same data but their forecasts are based on different information and therefore do not align. Reconciliation aligns all the forecasts on all levels by introducing minimal errors. As a result we end up with a single forecast aligned for all different subsets of demand. this is important to enable aligned decisions for capacity on both tactical and operational levels. Testing the proposed framework indicate that significant improvements are possible. 20% gains are achieved, in terms of the MASE, by the proposed framework compared to the current approach. By translating the results to yearly totals we find that our approach consistently forecasts the yearly total within 1% of the actual values. Reducing the error margin in 2017 by 89% from an overestimation of 9,3 FTE to 0,9 FTE. The increased accuracy allows better and more flexible decisions over capacity.
Item Type:Essay (Master)
KLM, Schiphol, Netherlands
Faculty:BMS: Behavioural, Management and Social Sciences
Subject:01 general works, 50 technical science in general, 54 computer science, 58 process technology
Programme:Industrial Engineering and Management MSc (60029)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page