In a canonical full-dataset benchmark evaluating wind-power forecasting, the gradient boosting model XGBoost achieved the lowest test mean absolute error of 0.0657, alongside a root mean squared error of 0.1134 and a coefficient of determination of 0.860. The research, conducted by Dionysios Alexopoulos, Christos Emexidis, and Kyriaki Saiti from New York College in Athens, Greece, was published in Frontiers in Energy Research. The investigation establishes a temporally partitioned, forecasting-first workflow utilizing the synthetic DaKS/Kassel wind-power dataset to address critical methodological challenges in grid-level renewable energy integration.
Accurate wind-power forecasting is essential for the reliable integration of renewable generation into electrical power grids, yet reproducible evaluation depends heavily on explicit temporal partitioning and clearly defined feature-availability assumptions.
The canonical tabular benchmark established a target-time-aligned baseline by combining DaKS day-ahead numerical weather prediction covariates with recent observed-power history constructed relative to target valid time. In this experimental design, all preprocessing parameters were fitted exclusively on the training partition rather than across the aggregate dataset.
To evaluate model feasibility across alternative model families, the researchers conducted controlled subset experiments across four individual wind parks. These comparative trials assessed several advanced neural and sequential architectures against the benchmark. The tested architectures included a standard multilayer perceptron, a gated recurrent unit network, a long short-term memory recurrent model, a temporal convolutional network, and PatchTST-Lite, a lightweight patch time series transformer. These subset experiments demonstrated the practical feasibility of deploying modern sequential neural network architectures within structured wind-energy forecasting pipelines.
Beyond standard point forecasting, the authors developed a specialized validation-calibrated residual diagnostic layer. This analytical layer was engineered to systematically examine prediction errors across three distinct operational facets: spatial distribution, temporal progression, and prevailing operating regimes.
When evaluated under the strong-model diagnostic configuration, the system achieved a test mean absolute error of 0.0662 while successfully retaining structured candidate warning episodes.
The authors emphasized an essential interpretive boundary regarding these diagnostic outputs, clarifying that the detected residual patterns represent structured deviations in forecast behavior rather than confirmed physical equipment faults. Consequently, this framework delivers a standardized, reproducible foundation for forecasting and residual screening in future wind-energy digital twins and prognostics and health management systems.
Subject of Research: Temporally partitioned wind power forecasting and residual diagnostics
Article Title: Temporally partitioned wind power forecasting with validation-calibrated residual diagnostics for PHM-oriented interpretation
Article References: Alexopoulos, D., Emexidis, C., & Saiti, K. (2026). Temporally partitioned wind power forecasting with validation-calibrated residual diagnostics for PHM-oriented interpretation. Frontiers in Energy Research, 14, 1948217. https://doi.org/10.3389/fenrg.2026.1948217
Image Credits: AI Generated
DOI: 10.3389/fenrg.2026.1948217
Keywords: wind power forecasting, temporal partitioning, XGBoost, residual diagnostics, prognostics and health management, digital twins, DaKS Kassel dataset, numerical weather prediction, multilayer perceptron, gated recurrent unit, long short-term memory, temporal convolutional network, PatchTST-Lite, mean absolute error, root mean squared error, renewable energy integration
News Source: Denise Maddox. (November 17, 2026). XGBoost Achieves Lowest Error in Temporally Partitioned Wind Power Forecasting Benchmark. Scienmag.



