Deep learning methods have been developed to improve decoding of brain activity and intention in EEG-based Brain-Computer Interfaces. The motor imagery paradigm is used for many purposes, one of which is neurorehabilitation to regain or improve motor control in patients after stroke. Improvements in the area shows that raw EEG signals can be used for decoding and classification with deep learning, but due to high latency and need for large amounts of training data, many methods are not feasible in real-time systems. In this thesis, 6 datasets of simulated EEG signals with two classes were generated, 3 with differences in frequency and 3 with differences in amplitude, and the ability of an Echo State Network of the reservoir computing paradigm to discriminate between the classes was investigated. The Echo State Network was also trained on different number of training samples with each dataset to examine how the classification accuracy changes with different number of training samples. The results show that Echo State Networks can achieve high classification accuracy and distinguish between signals, even when their features differ only slightly and the signal-to-noise ratio is low. Echo State Networks could potentially be used in real-time Brain-Computer Interfaces for neurorehabilitative purposes, but more research is needed with real EEG data.