This thesis examines the performance of computational intelligence in short term load forecasting. The main objective of the work was to propose and evaluate neural network, fuzzy logic, neurofuzzy and hybrid systems in the role of short term load forecast, considering some variables that affect the load behavior such as temperature, comfort indexes and consumption profile. The work consisted in four main steps: a study about load forecasting; the modeling of neural network systems, fuzzy logic and neurofuzzy related to load forecast; and case studies. In the load forecasting studies, some variables appeared to affect the behavior of the load curve in the case of electrical utilities. These variables include meteorological data like temperature, humidity, lightening, comfort indexes etc, and also information about the consumption profile of the utilities. It was also noted the distinct behavior of the load series related to the day of the week, the seasonableness and the correlation between the past and present values. A bibliographic research concerning the application of computational intelligence techniques in load forecasting was made. This research showed that neural network models have been largely employed. The fuzzy logic models have just started to be used recently. Neuro-fuzzy are very recent, and there are almost no references on it. The surveyed application projects using the three models were classified by its architecture, performance, measured errors, inputs considered and horizon of the forecast. In this work four systems were proposed and implemented for load forecasting: fuzzy logic, neural network, hierarchical neuro-fuzzy and hybrid neural/neuro- fuzzy. The systems were specialized for each day of the week, due to the different behavior of the load found for each of the days. For the neural network, neuro-fuzzy and hybrid, the data were separated in winter and summer, due to the different behavior of the load in each of the seasons. The fuzzy logic system was modeled for very short term forecasting using the historic load for each hour of the day, in steps of 10 minutes within each hour. The fuzzy system rules were generated automatically based on the historic load and the fuzzy sets were pre-defined. The system with neural network had its architecture defined through experiments using only load data, hour of the day and month as input. The network model chosen was the back- propagation. Tests were performed adding other inputs such as temperature and consumption profile. For the neural- fuzzy, a hierarchical neuro-fuzzy system, which defines automatically its structure and rules based on the historical data, was employed. In a further step, a hybrid neural/neuro-fuzzy was studied, so as the neural network forecast is the input for the neuro-fuzzy system. For the last three models, short term forecasting was made for one hour period. The proposed systems were tested in case studies, and the results were compared themselves and with results obtained in other projects in the same area. The load data of CEMIG between 1994 and 1996 was used in the fuzzy logic system in steps of 10 minutes for very short term forecasting. The performance was good compared with a neural network system using the same data. For the other models, short term load forecasting (I hour, 24 steps ahead) was done using the following data: load data of LIGHT and CPFL between 1996 and 1998; temperature (hourly for LIGHT and daily for CPFL); the codification of month and hour of the day; and a profile of load by consumption class. For doing. The error results obtained by the models were around 1,15% for the fuzzy logic, 2,0% for the neural network, 1,5% for the neuro-fuzzy system, and 2,0% for the hybrid system. This work has showed the applicability of the computational intelligence techniques on load forecasting, demonstrating that a preliminary study of the series and their relation with