The forecasting of reactive and active power is an important tool in the monitoring of an Electrical Energy System. The main purpose of the present work is the development of a new short-term reactive power hourly forecast technique, which can be used at utility or substations levels. The proposed model, named A Hybrid Model for Reactive Forecasting, is divided in two stages. In the first stage, the active and reactive power data are classified by an unsupervised neural network - the Self-Organized Maps of Kohonen (SOM). In the second stage, a Autoregressive Distributed Lags Model (ADL) is used with its parameters estimated by an Iteratively Reweighted Least Square (IRLS). It also includes a correction lag structure for serial autocorrelation of the residuals as used in the Cochrane-Orcutt formulation. The short term reactive power forecasting is divided in in sample and out of sample. The out of sample forecast is applied to hourly periods until one month ahead. The proposed model is applied to real data of one substation and the results are compared with two other approaches, a conventional Dynamic Regression and a Feedforward Multi-layer Perceptron (MLP) Artificial Neural Network model.