Short term electric load forecasting using an automated system of model choice

The paper studies the choosing mechanism of an energy company which gathers a library of electric load models and at every day chooses the best one for daily prediction. We use a combination of a semi-Markov process and a modified hidden Markov chain to describe the joint curve of loads, the daily best model, and exogenous information, of which temperature is an important factor. By extending the state space of the semi-Markov process, then in a computationally tractable way, the problem is embedded within a hidden Markov chain. Hence we can establish an EM algorithm and an enhancing statistical learning method for estimating parameters and forecasting load. Simulation reveals the range in which the proposed algorithm is applicable. Examples from real world datasets show that the proposed automated system is an alternate method for short term electric load forecasting with loads greater than a few hundreds MW. Supplementary material includes scripts of the proposed system and a guide of the scripts.