Decision tree-based approach for online management of fuel cells supplying residential loads

The paper demonstrates the online optimal management of PEM fuel cells for onsite energy production to supply residential loads. Classical optimization techniques are based on offline calculations and can not provide the necessary computational speed for online performance. In this paper a Decision Tree (DT) algorithm is employed to obtain the optimal, or quasi- optimal, settings of the fuel cell online and in a general framework. The main idea is to employ a classification technique, trained on a sufficient subset of data, to produce an estimate of the optimal setting without repeating the optimization process. The required training database is extracted by performing the optimization offline at different load demands as well as different natural gas and electricity tariffs using a genetic algorithm (GA). The approach provides the flexibility of adjusting the settings of the fuel cell online according to the observed variations in the tariffs and load demands. Results at different operating conditions are presented to confirm the high accuracy of the proposed generalization technique. In addition, the accuracy of the DTs to approximate the optimal performance of the fuel cell is compared to that of the Artificial Neural Networks (ANNs) used for the same purpose. The results show that the DTs can somewhat outperform the ANNs with certain pruning levels.