In modern complex and highly interconnected power systems, load forecasting is the first and most critical step in operational planning. The ability to predict load from few hours ahead to several days in the future can help utility operators to efficiently schedule and utilize power generation. The main focus of this research is to have an accurate and robust solution to the Short-term Load Forecasting (STLF) problem using Artificial Intelligence based techniques. Amongst several techniques reported in the literature, Artificial Neural Network (ANN) has been proposed as one of the promising solution for STLF. The ANN is more advantageous than statistical models, because it is able to model a multivariate problem without making complex dependency assumptions among input variables. By learning from training data, the ANN extracts the implicit nonlinear relationship among input variables. However, ANN-based STLF models use Backward Propagation (BP) algorithm for training, which does not ensure convergence and hangs in local minima more often. BP requires much longer time for training, which makes it difficult for real- time application. To overcome this problem, we use Particle Swarm Optimization (PSO) algorithm to evolve directly ANN by considering it as an optimization problem. With PSO responsible for training, we can modify ANN in any way to suit the problem or class of problems. Secondly, load series is complex and exhibit several level of seasonality due to which sometimes ANN is unable to capture the trend. To overcome this shortcoming, we have used modularized approach. We used smaller ANN models of STLF based on hourly load data and train them through the use of PSO algorithm. A variety of Swarm based ANN hourly load models have been trained and tested over real time data spread over a period of 10 years. Keeping in view the various seasonal effects and cyclical behavior, we divided the load data in different scenarios and results were analyzed and compared. The forecast results in majority of the cases are fairly accurate and prove the promise of proposed methodology. This approach gives better-trained models capable of performing well over time varying window and results in fairly accurate forecasts.