Streaming dynamic mode decomposition for short‐term forecasting in wind farms

Abstract

Forecasting in wind energy is a crucial task to perform adequate wind farm flow control or to participate in the energy market. While many power forecasting methods exist, it is notoriously difficult to capture both short‐ and long‐term variations in the wind farm system in real time. We demonstrate a data‐driven real‐time system identification approach to forecasting based on streaming dynamic mode decomposition methodology (sDMD). The method is capable of characterizing nonlinear, time‐varying, multidimensional time series data in a computationally efficient manner. The algorithm is modified to work with data streams by adjusting the dynamic mode decomposition continuously as new data are made available. The method is applied to high‐frequency SCADA data from the Lillgrund offshore wind farm. A 23.31% improvement over persistence forecasting is found for 5‐min‐ahead forecasts of the power output of all turbines in the wind farm. sDMD is shown to be a suitable tool for capturing short‐term dynamics while adapting to long‐term changes in wind speed and direction and has potential applications in real‐time wind farm control.

Towards fine tuning wake steering policies in the field: an imitation-based approach

“…Few experiments have so far evaluated RL methods on simulators with both dynamic wake propagation and turbulent wind inflow, but preliminary results also point towards decentralized learners being faster than centralized alternatives. In [27], an RL approach for wake steering where a centralized learner controls 3 turbines in a row of 4 was tested on HAWC2Farm [28] with a turbulence intensity of 7.5%. When the wind direction was aligned with the turbine row, which corresponds to the case with the largest wake effect, no significant increase in total power production was observed within the first 72h (ϕ ∞ = 2 • ) and 96h (ϕ ∞ = 0 • ) of the simulation.…”

Wind farm flow control: prospects and challenges

“…There are many other (quasi-)dynamic models, with different fidelities (representing the flow, possibly the turbines, etc.) that can take a similar role, such as FLORIdyn (Becker et al, 2022), FRED (Van Den Broek and van Wingerden, 2020), PossPOW (Göçmen et al, 2019), FastFarm (Jonkman and Shaler, 2021), HAWC2Farm (Liew et al, 2022), and the DWM model by Larsen et al (2008). Another modeling approach uses data from field measurements and/or highfidelity CFD to generate low-dimensional surrogate models.…”

“…Another modeling approach uses data from field measurements and/or highfidelity CFD to generate low-dimensional surrogate models. This approach, also referred to as data-driven modeling, uses techniques such as dynamic mode decomposition (DMD) to generate models (e.g., Liew et al, 2022;Chen et al, 2020;Annoni et al, 2016a;Cassamo and van Wingerden, 2020) and/or machine learning approaches (see also Sect. 3.3).…”

Development of a machine learning model for wind turbine fatigue and ultimate loads based on static loads

“…Multi-objective design and optimization tools are utilized for further pushing the design objectives, with design loads relying on full time-domain aeroelastic simulations [2], deterministic design cases [3,4,5], steady-state engineering approaches [6,7], or surrogate load models [8]. In this work, a machine learning model is trained on time-domain aeroelastic load simulations in order to provide a load surrogate model from static rotor loads input to lifetime wind turbine design loads [9,10], and evaluated on the IEA-3.4-130-RWT [11] with a range of design variations.…”