We proposed novel feature quantities of electroencephalogram (EEG) to effectively detect affects in humans. A machine learning model using the proposed feature quantities of time series EEG powers showed higher accuracy to estimate affective states of concentration and relaxation compared to a model using conventional EEG powers. Ten healthy human participants conducted a 3-back task with monetary reward to evoke a state of concentration and a 0-back task as relaxation, three times in different days. Their EEG signals from frontal areas were measured during each task period using a wearable devise. We first analyzed EEG powers' time series in theta and alpha frequency bands in shorter segmentations. The theta power was greater, and alpha power was smaller, statistically significant at most electrodes (p < .05), during the concentration task than during the relaxation task, certificated validity of our experimental manipulation to induce concentration and relaxation. We then proposed the novel feature quantities, the 2nd-order time series of EEG power (fluctuation of time series of time series of EEG power), as we found nontrivial fluctuation in time series of EEG powers during both tasks. An accuracy of estimation of two internal states by the machine learning model (Support vector machine) using proposed the 2nd-order EEG powers was outperformed by the model using conventional EEG powers (67.1% to 83.3%). These results suggest that feature quantities reflecting the meta-level pattern of fluctuations of EEG power should be beneficial to estimate affective states in humans.