Free Trial

Safari Books Online is a digital library providing on-demand subscription access to thousands of learning resources.

Use of Neural Networks for Modeling Energy Consumption bAckground Neural Networks (NN), also commonly referred to as Artificial Neural Networks, are information-pro- cessing models inspired by the way the densely interconnected, parallel structure of the brain processes information. In other words, neural networks are simplified mathematical models of biological neural networks. The key element of the NN is the novel structure of the information processing system. It is composed of a large number of highly interconnected processing elements that are analogous to neurons, and tied together with weighted connections that are analogous to synapses. NN are capable of finding internal representations of interrelations within raw data. NN are con- sidered to be intuitive because they learn by example rather than by following programmed rules. The ability to learn is one of the key aspects of NN. This typical characteristic, together with the simplicity of building and training NN, has encouraged their application to the task of prediction. Because of their inherent non-linearity, NN are able to identify the complex interactions between independent variables without the need for complex functional models to describe the relationships between dependent and independent variables. Recently, the NN approach has been proposed as a substitute for statistical approaches for classification and prediction problems. The advantages of NN over statistical methods include the ability to classify in the presence of nonlinear relationships and the ability to perform reasonably well using incomplete databases. The comparison of the results from NN and statistical approaches indicated that neural net- works offer an accurate alternative to classical methods such as multiple regression or autoregressive models (Feuston & Thurtell, 1994; AlFuhaid et al., 1997). Although the NN concept was first introduced in 1943 (McCulloh & Pitts, 1943), it was not used extensively until the mid-1980's owing to the lack of sophisticated algorithms for general applications, and its need for fast computing resources with large storage capacity. Since the 1980's, various NN architectures and algorithms were developed (e.g. the multi-layer perceptron (MLP) which is gener- ally trained with the error backpropagation algorithm, Hopfield Network, Kohonen Network, etc.). Consequently, NN models have been used extensively as a tool for modeling, control, forecasting, and optimization in many fields of engineering and sciences such as process control, manufacturing, nuclear engineering, and pattern recognition. use of nn in energy modeling In the area of energy modeling, the application of NN has been mainly limited to utility load forecast- ing. There are several hundred papers in the literature on the application of NN for utility hourly load forecasting. These clearly show the superior capability of NN models over conventional methods (such as regression analysis). Park et al. (1991) were among the first group of researchers to use MLP NN for hourly load forecast- ing. In 1992, Peng et al. used an improved NN that used an alternate formulation of the problem in which the input was mapped to the output by both linear and non-linear terms, and an improved method for selecting and scaling the input units. Kiartzis et al. (1995) used a MLP NN with 24 output neurons, one for each hour of the day (i.e. their model could forecast the next 24-hour load profile on an hourly basis). Chen et al. (1996) included humidity in their NN model in addition to ambient temperature to account for the effect of humidity on air-conditioning component of the load. In 1997, AlFuhaid et al. tested a cascaded artificial NN (CANN) to capture the sensitivity of the non-linear influence of temperature and 183