Handbook of air conditioning and refrigeration / Shan K

Artificial Neural Networks
ENERGY MANAGEMENT AND CONTROL SYSTEMS 5.51
● Diagnostics—assistance in identifying solutions to complex technical problems
● Design—assistance in the selection of the HVAC&R systems and subsystems
Basics. An artificial neural network (ANN) is massive interconnected, parallel processing,
dynamic system of interacting processing elements that are in some aspect similar to the human
brain. The fundamental processing element is called the neuron, which is analogous to the neural
cell in human brain. The neurons are set in layers, and thus a network is formed as shown in
Fig. 5.25. Inputs representing the variables that affect the output of the network are feeding forward
to each of the neurons in the following layers with an activation depending on their weighted sum.
Finally, an output can be calculated as a function of the weighted sum of the inputs and an additional
factor, the biases.
The ability to learn is one of the outstanding characteristics of an ANN. The weights of the
inputs are adjusted to produce a predicted output within specified errors. ANNs have been increasingly
used in recent years to predict or to improve nonlinear system performance in HVAC&R. An
ANN system is characterized by its net topology, neuron activations transfer, and learning method.
Net Topology. The structure of the network of an ANN, or net topology, depends on the data flow
mode, the number of layers, and the number of hidden neurons.
● In Miller and Seem (1991), there are two types of data flow modes: state and feed-forward models.
In the state models, all the neurons are connected to all the other neurons. In feed-forward
models, the neurons are connected between layers, as shown in Fig. 5.25a, and the information
flows from one layer to the next. Feed-forward models are the most popular and most often analyzed
models.
● In an ANN, there is always an input layer with the number of inputs equal to the number of parameters
(variables) that affect the output.
● There may be one or more hidden layers of neurons next to the input layer. The selection of number
of hidden layers and the number of neurons in each hidden layer remains an art. Curtiss et al.
(1996) noted that too many hidden layers and hidden neurons tend to memorize data rather than
learning. The hidden layers and hidden neurons must be sufficient to meet the requirement during
the learning process for more complex nonlinear systems. More hidden layers and hidden units
need more calculations and become a burden.
Among the 10 papers published from 1993 to 1996 in ASHRAE Transactions regarding developed
ANNs, most have only one hidden layer, some have two layers, and only one has three hidden
layers. None exceeds three hidden layers. Kawashima (1994) recommended that in an ANN only
one hidden layer is sufficient for load prediction.
● If the relationship between the inputs and output is more complex, i.e., nonlinear, and more inputs
are involved, then more neurons are needed in each hidden layer. Kawashima (1994) also suggested
that the number of neurons in each hidden layer exceed 2m � 1. Here m indicates the
number of inputs.
● There is always an output layer next to the hidden layer(s). It is preferable to have one neuron
(single output) in the output layer for simplicity. There may be two or more neurons for multiple
outputs.
Neuron Activation Transfer. In Miller and Seem (1991) and Curtiss et al. (1996), for each neuron
in the hidden and output layers:
1. The input activations to a neuron in the first hidden layer h, denoted by i 1n, can be