Linear and non-linear methods for prediction of peak load at University of São Paulo

• Using linear and non-linear models to peak load prediction.
• Using k-mean algorithm for data clustering.
• Choosing an appropriated model for load forecast of a university.
• The adaptive neuro-fuzzy inference system showed the better performance.

Predicting the peak load contributes to the enhancement of energy management. This paper presents some models to forecast the peak load of a campus of the University of São Paulo, aiming to choose the best one for generalization. The developed models were linear and non-linear, respectively, linear regression and Artificial Neural Networks and Adaptive Neural Networks Inference Systems. The data used was power demand, weather variables and calendar data, which were treated and normalized. For non-linear models, differentially of other researches, the goal of this work is the used of only one network for forecast model from the introduction of a new variable that differentiates the type of day, if is weekday or not. The results obtained reports a good coincidence between the predicted and real peak load for all developed models, however the accuracy is better for non-linear models, mainly for Adaptive Neural Networks Inference Systems.