Symbolic Regression applied to the Short-Term Load Forecasting on the Substation level

Abstract

Data modeling is an important problem in several areas of knowledge. Symbolic Regression is a technique that allows finding a mathematical relationship to describe a set of experimental data. Unlike traditional modeling methods, Genetic Programming allows us to find a mathematical expression that can be analyzed and interpreted. Multi-Expression Programming is a variant of Linear Genetic Programming, which has many advantages, making it suitable for real cases. In this work, we propose to apply this variant of Genetic Programming to discover models for load forecasting, one day ahead, using data fromtransformer substation located in the province of Tucumán, Argentina. First, multiple tests were performed using known Benchmark functions to analyze the algorithm's behavior and adjust parameters. We concluded that Multi-Expression Programming is adequate to find models in complex problems as short-term load forecasting, achieving a similar error level compared to other techniques.