Energy Consumption Forecasts by Gradient Boosting Regression Trees

Recent years have seen an increasing interest in developing robust, accurate and possibly fast forecasting methods for both energy production and consumption. Traditional approaches based on linear architectures are not able to fully model the relationships between variables, particularly when dealing with many features. We propose a Gradient-Boosting-Machine-based framework to forecast the demand of mixed customers of an energy dispatching company, aggregated according to their location within the seven Italian electricity market zones. The main challenge is to provide precise one-day-ahead predictions, despite the most recent data being two months old. This requires exogenous regressors, e.g., as historical features of part of the customers and air temperature, to be incorporated in the scheme and tailored to the specific case. Numerical simulations are conducted, resulting in a MAPE of 5-15% according to the market zone. The Gradient Boosting performs significantly better when compared to classical statistical models for time series, such as ARMA, unable to capture holidays.

Automated summary

In recent years, there has been a growing interest in developing accurate and efficient forecasting methods for energy production and consumption. Traditional linear approaches are insufficient in modeling the relationships between variables, especially when dealing with multiple features. This study proposes a Gradient-Boosting-Machine-based framework to forecast the demand of customers in different locations within the Italian electricity market. The main challenge is to provide precise one-day-ahead predictions using historical data that is two months old, which requires incorporating exogenous regressors and tailoring them to the specific case. Numerical simulations demonstrate that the Gradient Boosting method outperforms classical statistical models such as ARMA, particularly in capturing holidays.