Epilepsy is a neurological condition that makes people susceptible to seizures. A seizure is a change in sensation, awareness, or behavior brought about by a brief electrical disturbance in the brain. We have developed a novel approach to predict seizures. Assume that seizure occurrence follows a stochastic process with Poisson distribution. Wavelet transform is used to calculate the energy of a specific frequency band to remove noise in the signal and to pick up useful information. A dynamic model is developed to describe this process and a hidden variable is included in it. We assume that the initial state of hidden variable has Gaussian distribution and it follows the second order autoregressive (AR) process. The method of particle filters associated with neural networks is used to figure out the hidden variable. Four patients' intracranial EEG data are used to test our algorithm including 28 hours of ictal EEG with 14 seizures and 40 hours of normal EEG recordings. The minimum least square error algorithm is adaptively applied to the model in order to calculate the model parameters and one seizure from each patient is supposed to be known. The results show that our algorithm can successfully predict 9 out of the 10 seizures and average prediction time is 32 minutes before seizure onset. The sensitivity is 90% and the false prediction rate is approximately 0.1FP/h.