Spatiotemporal Wind Forecasting with Uncertainty
Transformer models that predict full probability distributions of wind direction — so wind-farm control can act on uncertainty, not just a best guess.
PROBLEM
Wake steering — deliberately misaligning upstream turbines to boost total wind-farm output — needs to anticipate wind direction minutes ahead. Point forecasts hide the risk: steering on a wrong guess costs more than not steering at all. Operators need to know not just the likely direction, but how confident to be.
APPROACH
I adapted TACTiS-2, a transformer architecture for multivariate probabilistic forecasting, to operational SCADA data from NREL’s AWAKEN wind-plant field campaign — benchmarked against Informer, Autoformer, and Spacetimeformer, all implemented in a companion model library on the GluonTS estimator API. The framework is model-agnostic and built for scale: PyTorch Lightning training, distributed Optuna hyperparameter searches on Slurm HPC clusters, and models that learn each signal’s behaviour and the dependencies between turbines, producing full predictive distributions rather than single values. An end-to-end pipeline feeds those distributions into the Wind Hybrid Open Controller, where FLORIS wake-steering simulations put a control value on every forecast.
SYSTEM
Three open-source repositories share the work: the training framework, the model library, and the controller that consumes the forecasts.
OUTCOME
Calibrated ultra-short-term forecasts with honest uncertainty bands, published as an open-source research framework and my master’s thesis — and the foundation of how I think about every forecasting problem since: a prediction without its uncertainty is a liability dressed as an answer.