Abstract: PNN is a mathematical interpolation method for seeking the smallest error between true value and predicted value through training sampled data. At first, the original seismic data volume is treated using log reconstruction?impedance inversion technology for improving the identification accuracy of the reservoirs. The filtering parameters is then selected using conjugate gradient algorithm to calculate the smallest test error and select the best convolution operator and number of attributes, allowing the seismic attributes and target parameters to keep the best match and avoid the overtraining of neural network. The case study by applying this technology to Y3 block has shown a good predicted effect and higher prediction precision compared with linear regression and traditional neural network methods