Page 182 - Neural Network Modeling and Identification of Dynamical Systems
P. 182

5.2 SEMIEMPIRICAL ANN-BASED MODEL DESIGN PROCESS           173

































                          FIGURE 5.2 Canonical form of the initial theoretical  FIGURE 5.3 Canonical form of the initial theoretical
                          model (5.5) discretized by the explicit Euler method.  model (5.6) discretized by the explicit Adams–Bashforth
                                                                       method.
                          is the inclusion of additional time delay ele-
                          ments for the ODE right hand side values at pre-  rows; thick arrows represent the connections
                          vious time instants. Canonical form of the recur-  with varying weights.
                          rent neural network derived from the discretiza-  Transition from a system of finite difference
                          tion of initial theoretical model (5.4)bythe Euler  equations to a neural network model represen-
                          method (5.5) and the Adams–Bashforth method  tation allows us to explicitly preserve the local-
                          (5.6) is shown in Figs. 5.2 and 5.3, respectively.  ization of function, inherent to the original con-
                            The following notation is used in Figs. 5.2–5.9:  tinuous time model as well as the correspond-
                          square-shaped elements represent external in-  ing discrete time approximation. This feature
                          puts (control variables), measured outputs, and  of semiempirical models allows us to “freeze”
                          time delay elements; round-shaped elements   specific parts of the model, i.e., to prohibit the
                          represent the neurons; neuron activation func-  modification of both its structure and the val-
                          tions are depicted schematically within the  ues of its parameters during the training phase.
                          round-shaped elements (for instance, in this ex-  This approach is justified for the models that
                          ample we use linear, quadratic, and sigmoidal  include some relationships derived from accu-
                          functions); arrows between the elements repre-  rate theoretical domain-specific knowledge, that
                          sent the corresponding input and output con-  is, there is little doubt concerning these parts of
                          nections; arrows without the input element rep-  the model. Other parts of the model that might
                          resent neuron biases; values of the biases and  be doubtful (i.e., these parts are suspected to
                          connection weights are labeled along the ar-  be the main reason of the low accuracy of the
   177   178   179   180   181   182   183   184   185   186   187