A novel hybrid physics-data driven encoder-decoder model for wind turbine wake predictions

Fast and accurate monitoring or prediction of wind turbine wakes is essential for real-time optimization of wind farm power generation. However, conventional approaches face significant challenges: direct field measurements using instruments like LiDAR are often economically infeasible due to high costs, high-fidelity numerical simulations are computationally prohibitive and slow for real-time applications, and purely data-driven machine learning methods typically suffer from poor physical interpretability and uncontrolled errors in full-field predictions. To address these limitations, we propose wakeEvoNet, a novel hybrid physics-data driven framework for real-time wind turbine wake monitoring/prediction. This approach strategically deploys sparse but high-fidelity measurement points in the near-wake region (within 3 rotor diameters downstream) to capture flow velocity profiles, which are then used as input for convolutional neural network (CNN) based encoder-decoder models to infer the downstream wake field (from 4D to 14D downstream). By feeding real-time monitored data at each time step, the model prevents temporal error accumulation common in sequential predictions. This hybrid methodology—combining direct measurements in the near-wake with neural network-based evolution for the far-wake—significantly reduces the number of required instruments while maintaining high accuracy across the entire wake field. In this study, we demonstrate the framework using the NREL 5 MW reference turbine under neutral atmospheric boundary layer conditions. High-fidelity computational fluid dynamics simulations employing large eddy simulation and the actuator line method are leveraged to generate the training datasets, and three convolutional neural network-based encoder-decoder architectures are developed: a benchmark U-Net (wakeEvoNet-1), a CBAM-enhanced U-Net (wakeEvoNet-2), and a novel CA-TA-UNet (wakeEvoNet-3) incorporating a physics-guided lightweight spatial attention mechanism tailored for wind turbine wake flows. The models utilize three consecutive upstream cross-sections to predict subsequent downstream fields, with results showing high accuracy (R2 scores consistently above 0.80) and robust temporal performance on test data, offering an efficient alternative to traditional methods for real-time wind turbine wake monitoring/prediction.

Subjects

Artificial neural networks

Hybrid physics-data method

Light-weight attention mechanisms

Wind turbine wake

DDC Class

621.3: Electrical Engineering, Electronic Engineering

620.1: Engineering Mechanics and Materials Science

Options

A novel hybrid physics-data driven encoder-decoder model for wind turbine wake predictions