Abstract

This thesis addresses the significant challenge of achieving precise demand prediction within the aviation aftermarket maintenance and spare parts management sector, particularly concerning intermittent parts. These components, characterized by irregular demand occurrences, present a formidable challenge due to the difficulty in accurately estimating their demand and setting appropriate stock levels. Historical approaches, relying on conventional demand forecasting techniques, often yielded inaccurate forecasts, resulting in slow inventory turnover and increased warehousing costs. To address this challenge, a broad spectrum of techniques was examined, ranging from traditional statistical models to modern machine learning and deep learning methods falling under the broader domain of artificial intelligence. Deep learning has garnered substantial attention in time series analysis for its exceptional forecasting performance. Real-world data from an aviation company was used to implement various forecasting models, including traditional methods like the exponential smoothing, and Croston, as well as machine learning models like SVR, Random Forest, and K-nearest neighbour. Deep learning techniques, including LSTM, GRU, and CNN, were prominently featured, with customized error metrics tailored to intermittent demand forecasting. The findings highlight that, on average, deep learning models, especially Gated CNN and LSTM, outperform other models and offer highly accurate forecasts for intermittent demand. This study serves as a reference point for choosing the most effective
forecasting method to support inventory planning in the aviation aftermarket, reducing costs, and enhancing service reliability. Moreover, its relevance extends to various industries dealing with
intermittent demand, offering valuable insights for improved demand forecasting.