Читать книгу Automation of Water Resource Recovery Facilities - Water Environment Federation - Страница 216

Artificial neural network models are nonlinear statistical data modeling tools loosely based on biological neural networks. They can model real world systems by tuning a set of parameters. These parameters, known as weights, describe a model that maps from a set of given inputs to an associated set of outputs. The process of tuning the weights to the best values is called training. In simple terms, an ANN is a data-driven empirical model.

Artificial neural networks come in several different structures that are each most suitable for specific types of tasks. Table 7.3 lists common types of ANNs and their applications.

Figure 7.6 shows the structure and components of a back propagation ANN model. Inputs are shown on the left of the figure and outputs are shown on the far right. The hidden layer is a sequence of nodes. The arrows between the inputs and the nodes each represent a model weight that is multiplied by the input value.

TABLE 7.2 Advantages and disadvantages of ANN models (Hill, 2010).

TABLE 7.3 Types of ANNs and their application areas.

Each node in the hidden layer and the output layer has a structure and functions as shown in Figure 7.7. The first function is a summation of the inputs multiplied by their respective weights. The activation function can take any form, but is most commonly monotonic. Typical activation functions include linear, hyperbolic tangent, sigmoid, step, and exponential.

FIGURE 7.6 Components of a back propagation ANN model.

FIGURE 7.7 Functions of a neural network node.

The weights of an ANN model must be calibrated. This process is called training and involves running an input and output data set through the model and incrementally improving the estimate of the weights. Training is computationally intensive, but is typically one of the easier steps in model identification. There are many software packages available with this capability.