A piecewise-linear regression was performed on the 1993 ASHRAE time-series data. For data set A, several preprocessing steps were done including normalisation, outlier removal, and trend removal. Then training was done via a straightforward step descent of the coefficients of piecewise-linear maps derived from representation theorems of piecewise-linear maps. For data set B, no preprocessing was done except for a straightforward normalisation. Again, a simple steep descent was used to adjust the coefficients of the maps. Despite its simplicity, the algorithm performs reasonably. This indicates that the classic linear algorithms have obvious limitations in solving real-world problems, so one should use algorithms incorporating nonlinearities of the problems.

KEYWORDS: algorithms, accuracy, calculating, electricity consumption, water consumption, temperature, humidity, competition