Computer systems that learn: classification and prediction methods from statistics, neural nets, machine learning, and expert systems
Learning internal representations by error propagation
Parallel distributed processing: explorations in the microstructure of cognition, vol. 1
| Hi-index | 0.00 |
This paper proposes an approach which combines a statistical method with Feedforward Neural Network to building a forecasting model using a limited amount of real historical data. The use of the forecasting model is to make one-step-ahead forecasts of a monthly water demand time-series. Two categories of experiments are conducted. The first is a uni-variate approach, using solely the water demand time-series. This approach is taken to construct two stand-alone forecasting models: a feedforward network and a statistical model. The second is a bi-variate approach, which involves noise filtering, data pre-processing based on the underlying components of the water demand series and use of an influential external parameter. This approach is taken to build a hybrid statistical and feedforward network model. The results of the experiments suggest that for the task of interpolation of series with a high noise-to-signal ratio, such as the random component of a time-series, we can easily build neural network models which incorporate parameters believed to be closely correlated, but difficult (or even impossible) to mathematically model how they are correlated. With the help of correlated parameters, neural networks manage to extract regularities from a seemingly random time-series due to the high noise-to-signal ratio. The ease in building mathematically complex causal relationships in neural networks continues to make it an appealling approach to building forecasting models.