Interconnected power systems are large, complex, often stressed and dynamic. Maintaining security of such systems requires frequent computation of accurate system limits. Due to continuous system changes, the limits determined off-line, based on a previously defined scenario, are unsuitable by the time they are applied. Thus, real-time limit derivation is necessary and must be done as fast and as often as possible. This requires reducing the system model, without compromising accuracy for areas outside the study zone. Ontario IESO (Independent Electricity System Operator) uses DSA (Dynamic Security Assessment) software for On-Line Limit Derivation (OLLD). This paper presents the model reduction algorithm developed by the IESO. The key feature of the algorithm is the use of a nonlinear optimization technique to determine model parameters of the equivalent machines. The algorithm runs PSS/E within MATLAB environment such that the dynamic data for a set of equivalent machines are tuned to minimize the difference between the transient responses obtained from the NERC full model and the corresponding responses obtained from the reduced model. The new reduced model provides system limits for Ontario that are practically same as original full model, speeds up limit calculation and reduces the volume of data. Extensive validations were done in the areas of stability, voltage, protection and tie-line loading, and a sample of results have been presented to establish its usefulness.