Accurate load forecasting becomes an important task for operating and planning of a power system to maintain the security of power supply dispatched to the consumers. This paper proposes an advanced method, namely Extreme Machine Learning (ELM), to forecast load in short time period. It is observed that implementation of the ELM can redu-ce cost and time significantly. Comparison results with a well known algorithm, called the Back Propagation (BP), show that the ELM can converge a hundred times faster than BP. In addition, the ELM needs 100 hidden neurons while the BP needs 2 hidden neurons to achieve similar result. This reveals that the number of hidden neurons is not a problem for ELM as long as there is sufficient memory to perform its computation.