Electroencephalogram (EEG) has been used in the domain of emotion recognition, especially during the experience from music stimulus. A number of works have been submitted with promising results in emotion prediction tasks. Unfortunately, the majority of literature did not sufficiently take into account a non-stationary characteristic of EEG signals which could differ in each recording session, and this issue might be underlying reason why such research could not be transferred into real-world application. In this paper, we are proposing a novel solution by introducing a method of normalization across session. In particular, we performed a comparison of several normalization techniques to explore various techniques to address the issue of non-stationary in EEG data. The three proposed techniques in this study are rescaling, z-score standardization, and frequency band percentage. In our experiment, we collected EEG data from ten subjects in two scenarios: consecutive session and time varied session. Our emotion prediction results suggested that z-score technique was superior to other normalization techniques based on using support vector machine (SVM). To encourage other researchers to test the efficiency of their own approach with multiple session data, our dataset is publicly provided.