Energy price classification in North Brazilian market using decision tree

This paper proposes a short-term energy price classification model using decision tree. The proposed model does not predict the exact value of future electricity price, but the class to which it belongs, established with respect to pre-specified threshold. This strategy is proposed since for some applications, the exact value of future prices is not required for the decision-making process. A feature selection technique is applied to identify the relevant attributes to the decision tree. Some decision tree algorithms are investigated such as CART and C5.0. These results are compared with NN, another classification technique commonly used. In addition, ensemble models combining CART, C5.0 and NN two by two are analyzed. The prediction time horizon is 1 week ahead and the system used is the Brazilian market, which adopts a cost-based centralized dispatch with unique characteristics of price behavior. The results show that an ensemble model of C5.0 and CART produced high accuracy of 99.3%, and can be an attractive tool to mitigate risks in purchasing power.