Technical Comparative Benchmarking Study: Advanced AI Hybrid Methods for Renewable Energy Farm Optimization and Forecasting

Majid Masoumi; Asghar Dashtiy; Mohammad Dehghan; Mina Rajabi

Source-linked AI summary

Technical Comparative Benchmarking Study: Advanced AI Hybrid Methods for Renewable Energy Farm Optimization and Forecasting

Majid Masoumi, Asghar Dashtiy, Mohammad Dehghan, Mina Rajabi

arXiv:2608.26613v1Aug 27, 2026cs.LG

TL;DR

Offshore renewable-energy applications require models that can handle structured layouts, spatial interactions, temporal dynamics, and noisy operational measurements. This study benchmarks diverse AI architectures across WEC and wind-farm datasets, finding that model effectiveness depends on problem structure, with RF–BiLSTM achieving the strongest overall forecasting error performance.

Problem
Offshore renewable-energy expansion involves structured layouts, spatial device relationships, temporal dependencies, and noisy real-world SCADA measurements that require comparative model evaluation.
Method
The study benchmarks conventional ML, ensembles, neural and recurrent networks, Transformers, graph models, and hybrid approaches across WEC layout data and Penmanshiel wind-farm SCADA measurements.
Results
Across tasks, tree ensembles performed strongest for structured WEC data, while RF–BiLSTM achieved the best overall forecasting error with MAE=150.5 kW and R = 0.94.
Takeaways & Limitations
No single AI architecture is universally superior; models should be matched to problem structure, with tree ensembles, graph networks, or hybrid recurrent models favored in different settings.
Takeaways & Limitations
The study does not perform comprehensive global optimisation of continuous device locations and relies on a limited set of candidate layouts, while robustness across sites and out-of-distribution datasets remains to be established.

Abstract

from arXiv · show

This study provides a comprehensive benchmarking of conventional machine learning (ML), ensemble learning, deep neural networks, recurrent architectures, Transformers, graph based models, and hybrid ensemble deep learning approaches under complementary renewable energy scenarios. Three datasets are considered: a large scale WEC dataset, a 16 WEC dataset, and operational 10 min SCADA measurements at the Penmanshiel wind farm. For structured WEC layout data, tree ensembles exhibited a clear advantage over conventional ML and neural predictors because randomized partitioning and boosting efficiently captured nonlinear layout power interactions without requiring explicit feature representation learning. The Extra Trees was the strongest model, achieving considerable results. Relative to the MLP baseline, this corresponds to an approximately 63.7% reduction in MAE, demonstrating the suitability of randomized tree ensembles for high dimensional structured WEC data. Also, STGCN reduced the MAE to approximately 167.0 kW and achieved R = 0.93 by explicitly learning spatial and temporal turbine interactions. The best overall forecasting accuracy was obtained by the RF BiLSTM hybrid, with an MAE=150.5 kW. Compared with standalone LSTM, this represents an approximately 75% reduction in MAE, while improving on STGCN by approximately 10.0%. Finally, the experiments reveal that no single AI architecture is universally optimal: randomized and boosted ensembles are particularly effective for structured WEC surrogate modeling, graph networks become advantageous when explicit spatial interactions dominate, and ensemble recurrent hybrids provide the strongest balance when nonlinear tabular relationships and temporal dynamics coexist.

1 Introduction

Offshore renewable-energy optimization and forecasting involve strongly coupled physical variables, nonlinear dynamics, and costly simulations, motivating comparative evaluation of methods and layouts. Existing WEC studies show that spacing, geometry, wave direction, and layout choice affect energy capture, but optimization is often limited to predefined configurations.

Optimization motivation: Offshore systems require optimization across site selection, farm size, device locations, spacing, infrastructure, control, and maintenance because these variables are strongly coupled.Changing one turbine position can alter downstream wake conditions, while offshore environments also involve waves, currents, turbulence, hydrodynamic interactions, and extreme weather.
Forecasting motivation: Accurate forecasting supports grid scheduling, reserve management, storage operation, trading, maintenance planning, and system control.Offshore forecasting is difficult because the underlying processes are nonlinear, multivariate, nonstationary, and spatially dependent.
Prior layout studies: Square and triangular WEC configurations show strong spacing effects on annual energy production in relatively large arrays.The study systematically compares these configurations and demonstrates that spacing materially affects array performance.
Research limitations: Prior optimization studies commonly restrict searches to predefined candidate configurations rather than comprehensive continuous device positioning.This boundary limits identification of globally optimal unrestricted layouts and can leave broader design-variable interactions unexplored.
Prior layout studies: Triangular arrangements can be favourable under multidirectional waves, whereas square layouts may perform well under more directional seas.This result highlights the dependence of favourable layout geometry on wave directionality and sea conditions.
Prior layout studies: Appropriately selected spacing can significantly improve array-level energy absorption in device-specific hydrodynamic optimization.Such optimization is limited in scalability because hydrodynamic calculations are expensive and dimensionality increases rapidly for larger farms.

2 Offshore Renewable Energy Data

The study evaluates optimization and forecasting methods across two WEC layout datasets and real Penmanshiel wind-farm SCADA data. These datasets span different farm sizes, environmental conditions, spatial search spaces, and operational forecasting challenges.

Dataset scope: Three datasets cover two WEC farm-layout problems and operational wind-power forecasting using real SCADA measurements.The combination supports comparisons under complementary renewable-energy conditions.
Large-scale WEC dataset: 63,600 samples with 149 numerical features represent candidate configurations of 49-WEC farms in the large-scale WEC dataset.The main target is total absorbed farm power, with individual-device outputs and interaction information providing hydrodynamic insight.
16-WEC dataset: 288,000 samples with 49 continuous features represent 16 fully submerged three-tether CETO WECs across Sydney, Adelaide, Perth, and Tasmania.The inputs describe converter x-y coordinates and individual-device power information, while the target is total absorbed farm power.
WEC dataset comparison: The two WEC datasets support scalability analysis by contrasting many observations across wave environments with a substantially larger spatial search space.Representative optimized Perth and Sydney layouts illustrate how placement influences hydrodynamic interactions and overall farm behavior.
Penmanshiel SCADA dataset: Penmanshiel contains 10-minute measurements from 2016 to mid-2021 for 14 Senvion MM82 turbines, excluding WT03 from the released data.Variables include environmental conditions, turbine operation, electrical measurements, and temperatures; historical multivariate SCADA data predict future turbine power.
Operational forecasting conditions: Real SCADA data introduce noise, missing observations, changing operating regimes, nonlinear responses, and cross-turbine spatial relationships.These properties make the dataset relevant for evaluating temporal, graph-based, and spatiotemporal forecasting methods.

3 Advanced AI Forecasting Models and Methodology

The benchmark evaluates diverse regression and ensemble architectures for renewable-energy prediction, emphasizing how ensemble diversity, boosting, and hybrid temporal modeling address nonlinear structured data and sequential dependencies. It also identifies conditions under which tree ensembles, graph models, or recurrent hybrids are most suitable.

Benchmark scope: The benchmark compares conventional regressors, advanced tree ensembles, neural models, and cooperative learning approaches for wave-farm power prediction.The evaluated methods include MLP, gradient-boosting variants, AdaBoost, CatBoost, Extra Trees, LightGBM, TabNet, and XGBoost.
Ensemble learning: Ensemble predictors combine multiple base learners whose training may be independent, sequential, or hierarchical, improving representation of nonlinear feature interactions.This is especially relevant to renewable-energy tabular data combining meteorological, operational, spatial, and device-level variables.
Bagging-based ensembles: Bagging reduces prediction variance through bootstrap-trained learners, while RF and Extra Trees add feature or split-threshold randomisation to increase learner diversity.The variance reduction is particularly useful for noisy measurements and heterogeneous operating conditions.
Boosting-based ensembles: Boosting constructs learners sequentially to reduce residual errors, with gradient boosting fitting new learners toward the negative gradient of a selected loss.XGBoost adds regularisation, while LightGBM, CatBoost, and AdaBoost use distinct efficiency or weighting mechanisms.
Why ensembles suit tabular data: Tree ensembles naturally capture thresholds, discontinuities, and high-order interactions without requiring explicit feature representation learning.Their inductive bias can balance flexibility and generalisation on finite-sized structured datasets, while generally requiring less scaling and architectural design than neural networks.
Hybrid ensemble–deep learning models: Hybrid ensemble–temporal models combine structured-feature processing with recurrent sequence modeling to predict future power at a specified forecasting horizon.The ensemble extracts nonlinear relationships, whereas the recurrent component captures temporal dependencies; RF–BiLSTM achieved the strongest overall error performance among the evaluated forecasting models.
Model-selection considerations: Ensemble effectiveness depends on sample size, feature structure, temporal organisation, noise, preprocessing, and hyperparameter configuration.Graph neural networks or Transformers may offer additional advantages when explicit spatial connectivity or long-range temporal dependencies dominate.

4 Experimental Results and Comparative Analysis

The benchmark compares diverse AI model families for WEC surrogate prediction and wind-power forecasting, showing that performance varies with data structure and that complementary evaluation criteria matter.

Benchmark design: The study selected models from recent wind- and wave-energy benchmarks and evaluated them in a unified comparative framework.The evaluated configurations are summarized for surrogate models and wind-power forecasting architectures.
WEC power prediction: Extra-TD achieved the lowest MAE and TDA among the compared models for 16-WEC farm power prediction.Leading ensemble models approached unity in R2 and achieved CCO values close to 1.
WEC power prediction: Advanced ensemble approaches maintained favorable MAE, R2, CCO, and TDA trade-offs in the 16-WEC comparison.The two-dimensional plots place desirable models toward regions combining low error with strong fit, agreement, and reliability.
Wind-power forecasting: STGCN achieved an MAE of approximately 167 kW, R = 0.936, and EVS = 0.876 for 10-minute-ahead wind-power forecasting.Its graph representation learns inter-turbine relationships jointly with temporal dynamics.
Wind-power forecasting: Hybrid ensemble–recurrent models outperformed standalone neural models by combining nonlinear feature interactions with temporal dependencies.CatBoostBiLSTM and HGBRBiLSTM achieved R-values around 0.94, while XGBoost-BiLSTM had greater metric variability.

5 Conclusions and Future Research Directions

The study benchmarks AI methods across WEC power prediction and short-term wind forecasting, finding that the best architecture depends on the underlying problem structure. Tree ensembles lead on structured WEC data, STGCN benefits spatial-temporal forecasting, and RF–BiLSTM delivers the strongest overall forecasting error.

Conclusions: Model effectiveness strongly depended on the structure of the underlying renewable-energy problem.The benchmark covered conventional ML, ensembles, neural, recurrent, Transformer, graph, and hybrid models across three complementary datasets.
Conclusions: Extra-TD achieved an MAE of 4.7 × 10^4 and R2 = 0.96 at the Perth site, reducing MAE by approximately 64% versus MLP.The result indicates strong performance on nonlinear, high-dimensional WEC layout and control variables.
Conclusions: STGCN achieved an MAE of 167 kW by explicitly representing inter-turbine relationships while learning temporal dynamics.This substantially outperformed standalone DNN and recurrent architectures.
Conclusions: RF–BiLSTM achieved an MAE of 150.5 kW and R = 0.94, reducing MAE by approximately 74% versus standalone LSTM and improving on STGCN by approximately 10.0%.The hybrid combines Random Forest nonlinear feature learning and variance reduction with BiLSTM temporal representation.
Conclusions: Model selection based on a single metric can be misleading because the highest correlation did not identify the most accurate or reliable predictor.Joint assessment of absolute error, correlation, explained variance, logarithmic error, and prediction stability was more informative.
Future research: No AI architecture was universally superior; future work should test robustness across additional sites, horizons, environmental regimes, and out-of-distribution datasets.The study also identifies uncertainty-aware forecasting, physics-informed learning, dynamic graphs, multimodal learning, and adaptive hybrids as future directions.

Supplementary Material

The supplementary material compares hybrid and evolutionary methods for WEC farm layout optimization, including their search strategies, constraints, evaluation settings, and practical limitations.

Supplementary comparison: Table S1 compares hybrid and advanced evolutionary methods for WEC farm layout optimization.The broader related method uses multiple population-based optimizers, surrogate modeling, and backtracking.
Constraints and limitations: Continuous WEC x–y locations are subject to farm-area and minimum-spacing constraints.The optimization is restricted to a small set of candidate configurations and cannot guarantee globally optimal continuous layouts.
Evaluation setting: The optimization targets total absorbed power and improved convergence under limited evaluation budgets.The WEC array is evaluated under six realistic Australian wave-energy sites.
Method characteristics: The cooperative optimization framework combines exploration, exploitation, surrogate assistance, and backtracking to improve convergence.Backtracking allows poorly positioned WECs to be reconsidered during optimization.
Constraints and limitations: The framework has increased implementation and tuning complexity, while experimental farm sizes remain relatively small compared with future commercial-scale arrays.These boundaries constrain direct extrapolation to larger commercial deployments.