The Neural network algorithm was adopted to the discrimination of the partial discharge patterns before and after the tree initiation from a needle-shaped void. Phase (φ), discharge amount (q) and discharge frequency (n) were measured for all discharge pulses in the partial discharge measurement period. φ-qand φ-q-n patterns before and after the tree initiation were learmed by neural network, using the back-propagation method. The network which learned φ-q-n patterns showed a good discrimination performance than that leaned φ-q patterns. The discrimination performance decreased when the input pattern was taken from the non-experienced sample. Stable discrimination was possible when the tree length exceeded the void length.