Please use this identifier to cite or link to this item:
https://www.um.edu.mt/library/oar/handle/123456789/92059
Title: | An event-based approach for resource levelling in IIOT applications |
Authors: | Azzopardi, Daniel (2021) (1) |
Keywords: | Internet of things Neural networks (Computer science) Data sets Machine learning |
Issue Date: | 2021 |
Citation: | Azzopardi, D. (2021). An event-based approach for resource levelling in IIOT applications (Bachelor’s dissertation). |
Abstract: | Resource management and optimisation have become a large part of the Industrial Internet of Things (IIoT). While the adoption of machine learning for forecasting resource requirements has allowed companies to increase efficiency, trim costs and optimise logistics, the field is still relatively new, leaving several approaches yet to be evaluated. Building on previous research, this study presents a priority-based tangible-resource allocation system specialised in mitigating the effects of foreseeable events encountered in day-to-day operations for predicting resource demands in IIoT-scale environments. The tool adopts a cascaded dual-model approach for forecasting allocations and refining errors using Bidirectional-Gated Recurrent Units (Bi-GRU) and Bidirectional-Long Short-Term Memory (Bi-LSTM) model combinations. An event handling component employing recency weighted averaging is incorporated into the cascaded models to enhance the prediction. Further components aimed at providing a complete priority-based system include limited-resource reallocation for working with predictions larger than the available resource pool and routing for proposing suitable strategies to move available resources between locations. To supply the necessary data required by the system, a data generation component is also proposed. Evaluated on IIoT-scale synthetic dataset instances, the proposed hyperparameter tuned models achieved a mean absolute percentage error (MAPE) of 6.97%-7.22%, with the Bi-LSTM Initial Prediction with Bi-GRU Error Correction (IPECBi-LSTMGRU) and Bi-GRU Initial Prediction with Bi-LSTM Error Correction (IPECBi-GRULSTM) models observed to achieve a lower average error than the fully LSTM and GRU-based counterparts. The use of event bias values on forecasts having preemptable events reported a significant average accuracy improvement of 2.89%. Furthermore, limited-resource reallocation proved essential for adjusting predictions despite resulting in a slight error increase. Finally, routing ensured a priority-based resource movement strategy across locations for negligible computation time while allowing relocation of resources as they become available. |
Description: | B.Sc. IT (Hons)(Melit.) |
URI: | https://www.um.edu.mt/library/oar/handle/123456789/92059 |
Appears in Collections: | Dissertations - FacICT - 2021 Dissertations - FacICTAI - 2021 |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
21BITAI009.pdf Restricted Access | 2.32 MB | Adobe PDF | View/Open Request a copy |
Items in OAR@UM are protected by copyright, with all rights reserved, unless otherwise indicated.