Dynamic Prediction Model Based On Process Neural Networks And Its Application

That Dynamic prediction is the problem exists in reality commonly in the project field and study of science. Be in applying, many systems are all not bad look upon as be to be changed into problem time-varied a kind of complicated nonlinearity, a little problem composes in reply knowledge since being short of a priori theory, and inside alternation and environmental factor coactions complexity, the precise being in progress describes and analyses the model very difficult to use the mechanism ascertaining that. For instance, the aerocraft engine function declines to forecast, chemical industry produce PID control; Oil field development produces energy forecasting. But in the reality applying, requiring to build a model and forecasting is nonlinearity development system, making use of the tradition neural networks major part to need to give the order fixing a static model in advance. Recently, the research building a model and forecasting about because of dynamic network, have represented neural networks building a model and forecasting new development direction.In the aspect of the development of the process neural network theory, one typical process neural network models with good adaptability to problem solving in practical engineering are proposed respectively in this dissertation from the point view of the connection style and the approximation capability, which are called nonhierarchical process neural network model. 3 types process neural networks with differient time aggregation mechanism whose input and output both are time-varying functions are constructed and their spatial-temporal aggregation operation and activation can reflect the space aggregation function of the time-varying input signals and the stage time cumulation effect in the input process at the same time.The three network are nonhierarchical neural network with time-varied input and output, nonhierarchical neural network with delayed, nonhierarchical neural network with stage time-varied input. The effectiveness of these four process neural network models with their corresponding learning algorithms are proved by the simulation tests.The condition prediction theory proposed in this dissertation mainly includes the Dynamic prediction theory based on the process neural network and process prediction theory based on the process neural network. At the same time, the time series prediction theory based on the process neural network is analyzed in this dissertation. From the point view of functional analysis, the time series short-term prediction can be seen as a functional approximation problem, and the time series long-term predictioncan be seen as an operator approximation problem. The functional approximation capability and the operator approximation capability of the process neural network are discussed and proved, which set the theoretic foundation for the time series prediction theory proposed in this dissertation. The effectiveness of the time serie prediction theory based on the process neural network is validated by the chaotic Mackey-Glass time series prediction. Dynamic prediction model based on process neural networks can meet nonlinear recognition and process predition of dynamic system, and has better adaptability to dynamic forecasting and prediction problem in mechanism. The paper analyzes the information transfer mechanism and theory property of the process neural networks, constructs a training error function based on general distance functional, gives 3 types network models learning algorithms based on function basis expansion integrated with gradient descent by time granularity segmentation, and proves the effectiveness of models and algorithms by the examples of Mackey-Glass chaotic time process prediction and power load forecasting.