A greedy algorithm for dimensionality reduction in polynomial regression to forecast the performance of a power plant condenser

Engineers and physicists agree in stating that the heat transfer rate and the cleanliness factor are good indicators of the thermal performance of a power plant condenser. Both, the heat transfer rate and the cleanliness factor depend on other physical variables such as mass flow rates, pressures and temperatures measured by several sensors in the power plant. There is no knowledge about the relationship beyond the intuition of the expert that it is polynomial, taking into account the flux and heat balances derived from the mass and energy conservation principles. Performing a common full polynomial regression of a certain degree-considering all possible monomials formed by the set of measured variables-leads to a NP-hard and NP-complete problem. Besides, it would be necessary to decide the degree of the polynomial beforehand. This paper overcomes this drawback by proposing a greedy algorithm based on polynomial stepwise regression. It previously selects relevant monomials from the set of possible ones and then performs a common polynomial regression with the relevant monomials. The method is tested on some artificial data sets and some UCI repository data sets before being applied to the power plant condenser. The results show that our method successfully outperforms other state-of-the-art learning methods, taking into account both effectiveness and efficiency.