| This thesis mainly studies the design method of discrete process neural network algorithm, and its application in reservoir identification.The input of process neural network is some time-varying continuous functions, which is not convenient to deal directly with discrete samples. To address this issue, First, a training algorithm of PNN based on piecewise linear interpolation function is proposed. Secondly, two training algorithms based on numerical integration are proposed. The cubic spline integration and the parabolic interpolation integration are used in the hidden layer to deal with the time-domain aggregation of discrete samples and weights. The classical neurons are used in output layer. In order to improve the convergence ability of the network, the Levenberg-Marquard algorithm is employed to adjust the networks’parameters. Thirdly, to enhance approximation ability and computation efficiency of multi-aggregation process neural networks (MAPNN), a training algorithm based on numerical integration is proposed. Fourth, an input sequence point by point mapping-based neural networks model, whose input of each dimension is a discrete sequence, is proposed. Fifth, a quantum neural network model, whose input of each dimension is a discrete sequence, is proposed.To enhance the training performance of discrete process neural networks, an improved quantum-behaved particle swarm optimization algorithm is proposed. The proposed method also uses quantum potential well as optimization mechanism, but a new method of establishing the center of the potential well is employed. In each of iteration, first, the fitness of each individual is calculated, and then the first K individuals with the greatest fitness are taken as a candidate set. Secondly, take an individual as the center of the Delta potential well through roulette select. By adjusting the other individuals move to the centre of potential well, a single-step optimization is completed. In the process of optimization, we make the K. value decreases monotonically to achieve a balance of exploration and exploitation. The proposed approach is applied to the extreme optimization of the parameters optimization of the quantum-inspired neural network, and the experimental results show that the proposed algorithm is obviously superior to the original one.To enrich discrete process neural network algorithm design theory, a discrete convolution process neural network model and algorithm are proposed. In the proposed algorithm, the time aggregation is achieved by the convolution of the input sequence and the weight sequence, which can avoid the error caused by the orthogonal basis expansion in general process neural network. The networks are trained by the gravitational search algorithm and L-M algorithm, which can improve the convergence capability of the network. Experimental results demonstrate the superiority of the proposed algorithm.To address the problem of reservoir identification, the identification scheme based on the discrete process neural networks is investigated. The main issues include the concept and classification of reservoir identification, the influence factors of reservoir categories, the identification method of the level of flooded layer in oilfield, and the identification method of some reservoirs with different types. Using the actual log interpretation database, the reservoir identification method based on discrete process neural network is introduced. The study provides a new way to the complex identification of reservoirs. |
