Load forecasting is evolving in a smart grid era from the days before personal computer (PC) and the PC era. Thanks to the wide-spread use of smart meters, more data with high resolution is available than ever before. Meanwhile, with development of computing techniques, new forecasting models that previously were not practical due to computing power constraints, have begun to be used for electric load forecasting. However, academia and industry still lack the benchmark accuracy for short term load forecasting. Additionally, industry is looking for more accurate short term load forecasts to operate the power system in a reliable manner and to more profitably trade energy. This dissertation proposed research into short term load forecasting (STLF) with both point and probabilistic outputs. We first conducted benchmark accuracy research with Tao’s Vanilla Benchmark (TVB) model. We then developed a STLF model with recency effect. Based on that, we developed sister models and sister forecasts. The sister forecasts had two main applications: (1) improve load forecasts via combining sister forecasts; and (2) generate accurate probabilistic load forecasts via quantile regression averaging sister forecasts. Last, we reduced the computation time of the forecasting process using techniques of high performance computing.Through case studies with Independent System Operator New England (ISONE), Global Energy Forecasting Competition 2012 (GEFCom2012) and Global Energy Forecasting Competition 2014 (GEFCom2014), we have contributed to the state-of-the-art from three aspects. By relieving some of the constraints of computing power, we showed that a recency effect model with inclusion of various quantitatively selected combinations of lagged and moving average temperature variables can help enhance the accuracy of load forecasting models. We also demonstrated that combining sister forecasts can further improve the forecast accuracy of recency effect models. Even simple average of the sister forecasts can outperform each individual forecast in the case studies used here. Additionally, we demonstrated that superior probabilistic forecasts can be generated by using quantile regression averaging sister load forecasts, compared with other benchmarks measured by both pinball score and Winkler score. Finally, computing time was significantly reduced using high performance computing techniques.