Utilizing a long short-term memory model for the prediction of refugee arrivals in Uganda
| dc.contributor.author | Abusa, Michael | |
| dc.date.accessioned | 2026-01-27T09:00:35Z | |
| dc.date.available | 2026-01-27T09:00:35Z | |
| dc.date.issued | 2025-11-26 | |
| dc.description | A dissertation submitted to the faculty of techno-science in partial fulfillment of the requirements for the award of the degree of master of science in artificial intelligence of Muni University | |
| dc.description.abstract | Uganda, currently hosting over 1.95 million refugees as of August 2025 (UNHCR/OPM), faces mounting challenges in forecasting refugee influxes, particularly from South Sudan, Sudan, and the Democratic Republic of Congo. Traditional estimation methods, such as registration, mass verification, and census, are reactive, resource-intensive, and lack predictive foresight. This study examines the application of Long Short-Term Memory (LSTM) deep learning models to predict refugee arrivals using time-series data from Rhino Camp Refugee Settlement, spanning the period from 2015 to 2024. Adopting a hybrid research design, the study modeled five influx scenarios, including emergency-only, protracted-only, and mixed, to evaluate the LSTM model’s predictive performance against a linear regression baseline. The RMSE, R2, and MSE metrics were used to assess the model accuracy. Testing the four hypotheses yielded the following insights: Ho1: The availability of reliable and high-quality time-series data enabled effective predictive modeling, leading to the rejection of Ho1. Ho2: The LSTM model demonstrated superior reliability and precision, especially under protracted influx conditions, thereby rejecting Ho2. Ho3: Significant differences emerged in model performance across scenarios. The protracted-only scenario achieved the highest accuracy (MSE = 0.0191, RMSE = 0.1276), while the emergency-only scenario recorded the highest error (MSE = 0.0251), confirming the rejection of Ho3. Ho4: A significant relationship was found between refugee arrivals and influx timing, particularly in structured displacement settings, leading to the rejection of Ho4. Overall, the LSTM model consistently outperformed the linear regression baseline, capturing complex temporal dependencies in refugee arrival patterns. Training times ranged from 6 to 20 minutes, underscoring computational efficiency. The study concludes that LSTM-based forecasting, when grounded in structured and temporally stable datasets, offers a scalable and ethically sound approach to enhancing early warning systems and humanitarian planning. These findings contribute to the growing body of knowledge on AI applications in displacement contexts and offer a framework for future research and operational deployment. | |
| dc.identifier.citation | Abusa, M. (2025). Utilizing a long short-term memory model for the prediction of refugee arrivals in Uganda (Unpublished graduate dissertation). Muni University, Arua, Uganda | |
| dc.identifier.uri | https://dir.muni.ac.ug/handle/20.500.12260/890 | |
| dc.language.iso | en | |
| dc.publisher | Muni University | |
| dc.subject | Memory model | |
| dc.subject | Refugee | |
| dc.subject | Uganda | |
| dc.title | Utilizing a long short-term memory model for the prediction of refugee arrivals in Uganda |