The Digital Cherenkov Viewing Device (DCVD) is an instrument used by authority inspectors to assess irradiated nuclear fuel assemblies in wet storage for the purpose of nuclear safeguards. Originally developed to verify the presence of fuel assemblies with long cooling times and low burnup, the DCVD accuracy is sufficient for partial defect verification, where one verifies that part of an assembly has not been diverted. Much of the recent research regarding the DCVD has been focused on improving its partial defect detection capabilities. The partial-defect analysis procedure currently used relies on comparisons between a predicted Cherenkov light intensity and the intensity measured with the DCVD. Enhanced prediction capabilities may thus lead to enhanced verification capabilities. Since the currently used prediction model is based on rudimentary correlations between the Cherenkov light intensity and the burnup and cooling time of the fuel assembly, there are reasons to develop alternative models taking more details into account to more accurately predict the Cherenkov light intensity. This work aims at increasing our understanding of the physical processes leading to the Cherenkov light production in irradiated nuclear fuel assemblies in water. This has been investigated through simulations, which in the future are planned to be complemented with measurements. The simulations performed reveal that the Cherenkov light production depends on fuel rod dimensions, source distribution in the rod and initial decay energy in a complex way, and that all these factors should be modelled to accurately predict the light intensity. The simulations also reveal that for long-cooled fuel, Y-90 beta-decays may contribute noticeably to the Cherenkov light intensity, a contribution which has not been considered before. A prediction model has been developed in this work taking fuel irradiation history, fuel geometry and Y-90 beta-decay into account. These predictions are more detailed than the predictions based on the currently used prediction model. The predictions with the new model can be done quickly enough that the method can be used in the field. The new model has been used during one verification campaign, and showed superior performance to the currently used prediction model. Using the currently used model for this verification, the difference between measured and predicted intensity had a standard deviation of 15.4% of the measured value, and using the new model this was reduced to 8.4%.