Being able to forecast cash demand is necessary in cash management.ATMs play a central role in cash management by ensuring publicaccess to cash, which is a legally protected right in Sweden. To ensurecash is available during all circumstances, it is essential to be able toforecast cash demand, both during normal circumstances and duringcrises, such as wars and pandemics. Forecasting cash demand undernormal conditions is challenging in itself, due to factors as temporalvariability, spatial variability between ATMs, the unpredictability ofindividual transactions, and the complexities of a real-world dataset.Anomalous events add a dimension of complexity to the forecastingproblem.We propose a solution to a real-world complex forecasting problemin a domain where state-of-the-art forecasting methods have not yetbeen explored. In order to improve forecasting results, we have utilized representation learning to cluster similar time series sequencestogether based on transactional patterns. Furthermore, we have implemented a novel approach to adapting the implemented forecasting models to anomalous events by altering learning techniques inthe embeddings of the forecasting models.We benchmark several state-of-the-art forecasting models for bothshort- and long-term prediction, and apply representation learning toembed and cluster ATM time series based on transactional similarity.Across all experiments, TimesNet generated the best SMAPE scoreof 21.72%. Further experiments show that both fine-tuning and contrastive learning enhance model robustness under anomalous conditions, with contrastive learning yielding the most consistent improvements.This thesis was conducted in collaboration with Bankomat AB.