Artificial Neural Networks And Its Applications

Artificial neural networks (ANNs) has a long development history and it has made great progress in recent years. It is quite adept at forecasting & prediction, pattern recognition and control application. More and more studies on ANNs-based hydrological modeling have been reported over the last decade. Large numbers of applications show that ANNs has so many advantages such as parallel-operation, distributed- memory, self-adaptability, error-permissibility and so on. So it is suitable for solving those hydrology problems that are hard to build model accurately and have high non-linear and many uncertain puzzles in hydrology & water resource. The key of ANNs is to build a theory-based and applicable network model. This is involved with three main aspects:(1) To optimize network algorithm,(2) To fix network structure reasonably,(3) To utilize information fully.According to above aspects, the paper achieved valuable applications combined with ANNs' applications in hydrology prediction and load forecasting. The major research work is outlined as follows.(1) About algorithm, the paper has presented a self-adaptive error BP algorithm which can prevent the networks from getting in the part least and can shorten the studying time. An exponential energy function has been used to modify weights in an ANNs model with sensitive ability. This model can expedite convergence rate. A temporal difference neural networks (TDNNs) model has been firstly employed for training weights. And ANNs have been combined with Kalman filter in two difference ways in the paper.(2) About network structure, the paper has firstly introduced a new network structurein hydrology that is named as recurrent neural networks with bias (RNNB). Two bias cells connect severally to hidden unit and output unit. Hidden unit receive information not only coming from input unit but also time lag of one step information from itself. This measure strengthens the memory of old input information. Because this structure can reflect the correlation of hydrology variable between anterior a period and posterior one, the RNNB model heightens ANNs' adaptability and enhances forecasting precision.(3) For find more data information, the paper presents two simple ways, which one is seeking-optimum detection method and the other is adding-weight detection method. In the ANNs model with sensitive ability a forgetting factor and an expecting factor varying with time are designed to utilize reasonably all kinds of information and then increase the model's practicability. Based on information uncertainty principle the paper has made qualitative analysis and qualitative calculation to the problem of over-fitting that is often meet in application of ANNs. The research had made in the paper contributes to analyze the puzzle of over-fitting.To keep a firm grasp on ANNs' application problems, network structure, model algorithm, and data information, the paper puts forward some pertinent improved ways. They improve ANNs' practicability in system simulation and calculation, especially in hydrology and water resource.