Power Consumption Forecast using Nonlinear Regression Methodologies

Abstract

In this paper, we present an innovating method to forecast time series. We focus on a long-term forecasting model dealing with main periodicities in addition to short term effects. The consolidation of Fourier Transformations, which covers the basic oscillation of a time series, with machine learning algorithms approximating the error term, is at the heart of our forecasting model. Later on, we compare different regressions such as kernel and logistic regression, a combinatorial technique on sparse grids and finally a special representative of neural networks. Thereby we are able to forecast power consumption in certain locations of a given network and we show the results of those forecasts as functions of various inputs. The results presented are used for power demand planning of cities and are consequently prognostic in nature. In the context of Health Management, however, one usually works with anomaly detection and supervised learning methods. Nevertheless, a time series forecast in neighboring applications, e.g. the power consumption of a traction system in railway vehicles, could substantially benefit from these prognosis functionalities. This also means that deviations of physical quantities measured under real-time conditions from their expected behavior indicate a likely prevailing malfunction.