A preprint reports that an automated way of building inputs for energy forecasts was associated with an average 83.22% reduction in normalized root mean squared error, a measure of forecast error, compared with using no feature engineering. The result came from 18 test sets, but the improvement varied widely, from 28.22% on the weakest result to 98.39% on the strongest.
The system, called AutoEnergy, creates interpretable features from timestamps and past energy use through rule-based transformations, then integrates those features with automated machine-learning tools. Its stated aim was to reduce reliance on domain knowledge in energy-consumption forecasting and to support downstream optimization across diverse energy systems.
What the benchmark measured
For the main benchmark, each dataset allocated 80% of its data to training and 20% to testing. The comparison used normalized error and nonparametric statistical tests, with a significance threshold of 0.05, to assess differences among the feature-engineering approaches.
Across the benchmark, AutoEnergy's mean normalized error was 0.0338. Against the other tested approaches, labelled FT, TSMin and TSEff, the reported error reductions were 19.52%, 82.38% and 77.98%, respectively; the reduction versus the no-feature baseline was 84.72%.
The benchmark also recorded a mean processing time of 471.94 seconds, reported as 1.31 times faster than FT and 4.41 times faster than TSEff. The wide spread in error reduction makes the result dataset-dependent rather than a single expected gain for every setting.
From forecasts to battery decisions
The study then moved from forecasting accuracy to battery energy-storage decisions. Its BESS evaluation used a 55-day real-world electricity-demand and battery dataset from a single property in the UK, with approximately half the data, 27 days, for training, 14 days for validation and 14 days for testing. It compared three ANN-based methods, PTO, SPO+ and DBB, across ten independent runs with different random seeds.
The BESS comparison used normalized regret, the study's score for how far a decision fell short in the test. On the reported test horizon, SPO+ had mean regret of 0.0672, compared with 0.2046 for PTO with AFE, the paper's automated feature-engineering setup, an approximately 67.16% lower figure.
The abstract also reported 22.9% to 56.5% lower operating costs for models combining AFE with decision-focused learning than for the same models without AFE. The report gives those changes as percentages, not as underlying absolute costs.
A result with a narrow boundary
That economic result is not the same as a settled AFE effect. Reported regret gains from AFE were 56.48% for SPO+, 19.42% for PTO and 22.87% for DBB, but within-method comparisons did not remain statistically significant after Bonferroni adjustment. A separate global comparison did report significant differences among the BESS methods, with a p-value below .001, and ranked SPO+ with AFE best by regret.
The boundary around the battery findings is clear. They come from a 55-day winter case study at a single property in the UK and do not establish that the same cost or regret improvements would hold under other regions, tariffs, market structures or regulations. The forecasting result also varied substantially across the 18 test sets, so neither part of the analysis should be read as a universal guarantee.
Paper data and sources
Original title: Automated Feature Engineering, AutoML, and Decision-Focused Learning for Improved Energy Consumption Forecasting
Authors: Nasser Alkhulaifi
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text