Multi-horizon wind speed prediction with physics-embedded state space neural operator learning

Accurate and repeatable long-range wind-speed prediction is required for the integration of renewable energy into power grids, and these models will need to reconcile a number of contradictory requirements: high accuracy, low computational effort, physical plausibility, and quantified prediction uncertainty. In this paper, Physics-Embedded State-Space Operator Learning (PESO-Net) is proposed for multi-site, multi-horizon forecasting. The framework comprises a selective state-space encoder, gated cross-site mixing, a validation-controlled vertical shear regularizer, direct quantile decoding, and chronological conformal calibration. Observations from five years of 1-h synchronization at four anonymous sites were split chronologically into training, validation, calibration, and test sets, and one 168-h forecast trajectory was assessed over the 1–24 h, 1–72 h, and 1–168 h prefixes. The primary validation-weighted PESO-Net achieved MAE 2.080 m/s, RMSE 2.622 m/s, R² 0.176, 95.56% empirical coverage, 10.489 m/s mean interval width, and interval score 12.048. The primary validation-weighted PESO-Net had lower observed point errors and interval scores than the compared baselines. Among the four added model variants, PatchTST was closest by RMSE (2.727 m/s), with an interval score of 12.513. Mean predicted shear exponents ranged from 0.229 to 0.235, compared with observed means near 0.222; correlations of 0.287–0.302 support only limited physical agreement. The primary paired block-bootstrap RMSE contrast with Transformer was − 0.0185 m/s (95% CI − 0.0390 to 0.0042; p = 0.096), so statistical superiority was not established. The test period was used in earlier experiments, which limits the validity of claims of independent generalization. Thus, PESO-Net is not necessarily state-of-the-art for the four sites studied, but according to the data, it is competitive.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-15
DOI
https://doi.org/10.1038/s41598-026-80497-y
Primary Topic
Energy Load and Power Forecasting
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Multi-horizon wind speed prediction with physics-embedded state space neural operator learning

Karna Vishnu Vardhana Reddy, Srikanth Parikibandla, Alluri Sreenivas, Chiranjeevi Nuthalapati et al.
Scientific Reports
Energy Load and Power Forecasting
article

Multi-horizon wind speed prediction with physics-embedded state space neural operator learning

Karna Vishnu Vardhana Reddy, Srikanth Parikibandla, Alluri Sreenivas, Chiranjeevi Nuthalapati, C P Pavan Kumar Hota
article en

Abstract

Accurate and repeatable long-range wind-speed prediction is required for the integration of renewable energy into power grids, and these models will need to reconcile a number of contradictory requirements: high accuracy, low computational effort, physical plausibility, and quantified prediction uncertainty. In this paper, Physics-Embedded State-Space Operator Learning (PESO-Net) is proposed for multi-site, multi-horizon forecasting. The framework comprises a selective state-space encoder, gated cross-site mixing, a validation-controlled vertical shear regularizer, direct quantile decoding, and chronological conformal calibration. Observations from five years of 1-h synchronization at four anonymous sites were split chronologically into training, validation, calibration, and test sets, and one 168-h forecast trajectory was assessed over the 1–24 h, 1–72 h, and 1–168 h prefixes. The primary validation-weighted PESO-Net achieved MAE 2.080 m/s, RMSE 2.622 m/s, R² 0.176, 95.56% empirical coverage, 10.489 m/s mean interval width, and interval score 12.048. The primary validation-weighted PESO-Net had lower observed point errors and interval scores than the compared baselines. Among the four added model variants, PatchTST was closest by RMSE (2.727 m/s), with an interval score of 12.513. Mean predicted shear exponents ranged from 0.229 to 0.235, compared with observed means near 0.222; correlations of 0.287–0.302 support only limited physical agreement. The primary paired block-bootstrap RMSE contrast with Transformer was − 0.0185 m/s (95% CI − 0.0390 to 0.0042; p = 0.096), so statistical superiority was not established. The test period was used in earlier experiments, which limits the validity of claims of independent generalization. Thus, PESO-Net is not necessarily state-of-the-art for the four sites studied, but according to the data, it is competitive.

Scientific Reports
Indian Institute of Technology Guwahati (IN), SRM University, Andhra Pradesh (IN), Koneru Lakshmaiah Education Foundation (IN), GITAM University (IN)
Energy Load and Power Forecasting
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon