Abstract:

When applying the feed-forward network to describe the nonlinear relationship between seismic attributes and reservoir parameters, the classical error back propagation（BP）learning algorithm has slow convergence and
is easy to fall into local minima, and has many other deficiencies. This paper researches a new reservoir parameter prediction technique by seismic multi-attribute which is based on the integration of particle swarm optimization
（PSO）algorithm and BP algorithm of the feed-forward network. PSO is agroup random search evolutionary computation technique, has a faster convergence rate and strong global search ability; BP algorithm is essentially a gradient descent algorithm, focusing on local search. Hybrid algorithm trains network by PSO and BP alternatively. That is, when the learning error energy of PSO in required generation has not changed, using BP for local search disturbances. Function approximation test and practical reservoir parameter prediction experiment illustrate the hybrid learning method has a short training time, high efficiency and reliability, and has good prospect.

Key words: error back-propagation algorithm, particle swarm optimization, feed-forward neural network, seismic attribute, porosity