Each student log to the learning management system (LMS) generates data on who accessed the LMS, when, from which IP address, and which activity is performed. It is the way to collect the large amount of data on students and not to disturb the process of learning while the data are collected. The available data could be also used in the way that student logs to a certain course on the LMS are regarded as time series. In this paper it is proven that the observed time series carry the information on student behavior which is indicative for the final exam success. By applying the Fourier transform (FT) to the created time series, i.e. by switching time series to the frequency domain, the opportunity occurs to reduce the amount of data and for the data processing to become more operational. The second reason to apply FT is to reveal the hidden periodicities which are the key to modeling and prediction, and the third reason is to reduce the noise and to achieve better accuracy of student success prediction. The hypotheses of this study are tested on data sets on students enrolled in two schools, in four different courses, and during two academic years. The sample consists of data on approximately 1,400 students. This research has demonstrated that it is possible to create the student success prediction model through data mining on blended learning environment data using frequency domain time series analysis of student logs to LMS. The proposed model achieves the prediction accuracy which is equal or bigger of the prediction accuracies achieved by similar models described in the literature. The technique for compressing time series of student logs to LMS is developed. The technique removes the noise, meaning that using the reduced number of coefficients gives better prediction accuracy than using models based on the full amount of data. Furthermore, it is proven that it is possible to achieve the success prediction with the accuracy more than 75% based on data for the first 2/3 of the total length of the course. The fundamental contribution of the paper is the application of spectral analysis, i.e. analysis of time series in the frequency domain to the data set from the hybrid learning environment.