The last three decades have experienced an impressive growth in loads interfacing with the grid through power electronic devices. These include personal computers and most office equipment as well as industrial induction motors driven by variable frequency drives. The increasing importance of these nonconventional loads has an impact which is yet to be quantified and modelled systematically. This is particularly the case with the natural load-side frequency response. This lack of knowledge compounds with the current uncertainty in the characterization of the load's natural behaviour under frequency disturbances.The load in power systems is known to have an inherent frequency sensitivity. Because of the complexity and the constant evolution of the aggregate system load, the frequency sensitivity is difficult to estimate. In fact, in practice, utilities use some ill-defined, static and generic values for this sensitivity as part of their dynamic performance assessments and load-frequency control designs. To address this shortcoming, this thesis proposes a methodology based on a bottom-up approach based on aggregate load composition information and load behaviour modelling. The method allows for the development of a deeper understanding of the currently unknown impact of the load mix and other external factors on the natural load-side frequency sensitivity. The thesis then goes on to demonstrate that the load-frequency sensitivity is indeed a time-varying value, changing throughout the course of the day and the year, unlike the industry-wide static value assumption. The methodology allows for the identification of the explanatory factors behind the sensitivity variations. The evidence shows that those explanatory factors are the outside ambient temperature, the time of day and the date. The methodology is also able to model frequency sensitivity within a 10% error margin (using only publicly-available data).A comprehensive sensitivity analysis is undertaken to prove the robustness of the Load Frequency Sensitivity Bottom-Up Methodology (LFS-BUM). Probabilistic characterizations that represent the uncertainty embedded in the load-frequency sensitivity (LFS) at various times reveal that generally the prediction error probability distributions are skewed and sometimes even bimodal. For this reason, Edgeworth series expansions are used to provide analytical formulae for the approximate the probability distribution of the LFS values. The statistical relationship between LFS and the outside ambient temperature is used to generate prediction models for LFS. Error analyses of the models demonstrate an ability to predict LFS within a 7.66% error margin. The methodologies and the results should be of interest most notably to transmission system operators as they attempt to quantify the resources necessary to conduct primary frequency control tasks reliably and at the best possible cost.Furthermore, the bottom-up methodology is used to provide a detailed understanding of the evolution of LFS and thus is of potential use in informing grid operators and planners on the potential impacts of an evolving load mix which is increasingly insensitive/"stiffer" to frequency deviations, especially for frequency control. In addition, it will assist as a primary tool in redesigning frequency regulation function to accommodate for load participation.