This paper presents a new methodology for electricity demand forecasting on very short-term horizon based on a discrete probabilistic model (Markov Chain). The modeling process is automated by a feature extraction tool, the Self-Organizing Map, considering historical data of climate variables (air temperature, relative humidity and wind speed) and load behavior, related through the thermal discomfort index and wind chill. Thus, it is possible to estimate the probability of a certain demand level occur given a current climatic condition, as well as the number of time intervals (hours) until this occurs. The forecast is then used to control the decentralized dispatch of a small hydroelectric power plant, aiming to minimize overtaking the transmission contract.