<p>Energy price forecasting models are crucial decision-support tools for energy trading planning, mainly in emerging markets with a liberalization agenda. Machine learning (ML) forecasting models have recently surpassed traditional statistical models, mainly due to their inherent complexity and trivial capacity to leverage external sources of information. We propose a financial return-weighted ensemble of ML-based models to forecast energy products, where each committee member specializes in different data sources that influence energy prices. We also introduce feature set enrichment and feature selection pipelines to select the best variables to train the Gradient-boosted Decision Trees that compose our solution. We name our proposal Weighted Electricity Ensemble Forecaster based on Feature selection and Multi-source data (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\text {W}[\text {EF}]^2\text {M}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>W</mtext> <msup> <mrow> <mo stretchy="false">[</mo> <mtext>EF</mtext> <mo stretchy="false">]</mo> </mrow> <mn>2</mn> </msup> <mtext>M</mtext> </mrow> </math></EquationSource> </InlineEquation>) and showcase an instance of it in the Brazilian free energy market, where we forecast energy price contracts with liquidation in one month in the future. These contracts have the most liquidity in the Brazilian free energy market and are highly influenced by climate variables, economic factors, and changes driven by the decisions taken in the Brazilian Electricity System’s operation planning. Our proposed ensemble of ML forecasters consistently surpasses traditional statistical-based forecasting models, other ML solutions, and a strong baseline when forecasting closing prices 1, 5, and 21 business days ahead. In our realistic simulation studies, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\text {W}[\text {EF}]^2\text {M}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>W</mtext> <msup> <mrow> <mo stretchy="false">[</mo> <mtext>EF</mtext> <mo stretchy="false">]</mo> </mrow> <mn>2</mn> </msup> <mtext>M</mtext> </mrow> </math></EquationSource> </InlineEquation> provides the smallest error values and highest financial returns. Namely, our proposal improves the returns over the second best-compared solution at around <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$17.86\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>17.86</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> for 1 business day, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$18.14\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>18.14</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> for 21 business days, and 7 times for 5 business days ahead.</p>

Short-term energy market forecasting ensemble with multi-data source integration

期刊论文
发表日期：

2025年11月24日

查看全文

Saulo Martiello Mastelini,
Marcos Basile Saviano de Paula,
Moisés Rocha dos Santos,
Sérgio Baldo Júnior,
Lucas Menezes Ladeira,
Luciano Contin Gomes Leite,
Ronan Gustavo Carvalho Furtado,
Thiago Felicio de Souza,
Ewerton Guarnier,
Donato da Silva Filho

摘要

Energy price forecasting models are crucial decision-support tools for energy trading planning, mainly in emerging markets with a liberalization agenda. Machine learning (ML) forecasting models have recently surpassed traditional statistical models, mainly due to their inherent complexity and trivial capacity to leverage external sources of information. We propose a financial return-weighted ensemble of ML-based models to forecast energy products, where each committee member specializes in different data sources that influence energy prices. We also introduce feature set enrichment and feature selection pipelines to select the best variables to train the Gradient-boosted Decision Trees that compose our solution. We name our proposal Weighted Electricity Ensemble Forecaster based on Feature selection and Multi-source data ( $\text {W}[\text {EF}]^2\text {M}$ $W {[EF]}^{2} M$ ) and showcase an instance of it in the Brazilian free energy market, where we forecast energy price contracts with liquidation in one month in the future. These contracts have the most liquidity in the Brazilian free energy market and are highly influenced by climate variables, economic factors, and changes driven by the decisions taken in the Brazilian Electricity System’s operation planning. Our proposed ensemble of ML forecasters consistently surpasses traditional statistical-based forecasting models, other ML solutions, and a strong baseline when forecasting closing prices 1, 5, and 21 business days ahead. In our realistic simulation studies, $\text {W}[\text {EF}]^2\text {M}$ $W {[EF]}^{2} M$ provides the smallest error values and highest financial returns. Namely, our proposal improves the returns over the second best-compared solution at around $17.86\%$ $17.86 %$ for 1 business day, $18.14\%$ $18.14 %$ for 21 business days, and 7 times for 5 business days ahead.