Combining forecasts is a well-known approach to further improving forecast accuracy. In the load forecasting literature, there are only few papers discussing load forecast combination. Most of them are on combining independent forecasts. However, in practice, load forecasters may be able to concentrate on only one or few particular forecasting techniques due to limitations in educational background, time for model development, costs of additional software and so forth. How to conduct forecast combination with these real-world constraints is a challenging problem. This thesis proposes a novel solution to the aforementioned problem by combining sister load forecasts, which are generated from a family of sister models sharing very similar model structure developed from similar variable selection processes. In this thesis, 13 forecast combination methods are tested on four sister models. Through a comprehensive case study using publicly available data from the Global Energy Forecasting Competition 2014, combining sister forecasts using simple methods is found to outperform each individual forecast. In addition, the regression based combination, which uses a regression model to combine sister forecasts, outperforms the other methods for the aforementioned data set. Comparing with the best individual model, the regression-based combination reduces the forecast mean absolute percentage error (MAPE) by approximately 9%. It also outperforms simple average by 11 %. Note that simple average may not always outperform the best individual forecast, which is shown in this test case.This thesis starts the first formal investigation on combining sister forecasts, which is shown to be very effective in improving load forecasting accuracy of individual models. The proposed approach is of great practical value in the sense that it leverages existing variable selection processes and does not require additional skill sets from the load forecaster.