Model-free estimation of available power using deep learning

Abstract

Abstract. In order to assess the level of power reserves during down-regulation, the available power of a wind turbine needs to be estimated. The current practice in available power estimation is heavily dependent on the pre-defined performance parameters of the turbine and the curtailment strategy followed. This paper proposes a single-input model-free approach dynamic estimation of the available power using recurrent neural networks. Accordingly, it combines wind turbine control considerations and modern forecasting methodologies for a model-free, single-input estimation of available power. It enables a robust real-time implementation of dynamic delta control, as well as higher-accuracy provision of the reserves to the system operators. The model-free approach requires only 1 Hz wind speed measurements as input and estimates 1 Hz available power as output. The neural network is trained, tested and validated using the DTU 10 MW reference wind turbine HAWC2 model under realistic atmospheric conditions. The unsteady patterns in the turbulent flow are represented via long short-term memory (LSTM) neurons which are trained during a period of normal operation. The adaptability of the network to changing inflow conditions is ensured via transfer learning, where the last LSTM layer is updated using new measurements. It is seen that the sensitivity of the networks to changing wind speed is much higher than that of turbulence, and the updates are to be implemented solely based on the altering inflow velocity. The validation of the trained LSTM networks on time series with 7, 9 and 11 m s−1 mean wind speeds demonstrates high accuracy (less than 1 % bias) and capability of transfer-learning online. Including highly turbulent inflow cases, the networks have shown to comply with the most recent grid codes, which require the quality of the available power estimations to be evaluated with high accuracy (less than 3.3 % standard deviation of the error around zero bias) at 1 min intervals.

Virtual sensors for wind turbines with machine learning‐based time series models

“…Given the goals of the current study, it is most important to understand how to treat the inputs in order to achieve good model performance. Therefore, we carry out a systematic tuning on the second type (data‐related) hyperparameters using a grid search, while for the first group of hyperparameters, we choose the appropriate values (shown in Table 3) based on experience from previous studies such as Göçmen et al and Schröder et al 19,28 …”

“…For example, a notable exception from the aforementioned research gaps is the problem of power prediction, where the use of data-driven surrogate models with time dependence included is common. Here, the state-of-the-art deep learning algorithms are increasingly adopted, among which recurrent neural networks, such as Long-Short-Term Memory (LSTM), stand out as promising approaches for time-dependent dynamic problems, for example, 19,20 based on simulations and large-scale field measurements, respectively. With the help of a fixed but learned mapping function, LSTMs combine the input vector with their state vector and are particularly designed for sequence prediction problems.…”

Probabilistic surrogates for flow control using combined control strategies

“…The behaviour of the resulting Weibull CDFs with respect to the given control settings are mapped via polynomial chaos expansion (PCE) approach [18,19]. Given their lower computational cost (than e.g., Kriging [19]) as well as higher interpretibility, explainability and reliability under limited data (compared to e.g., neural networks [20,21]), PCE is considered to be the most suitable choice for the surrogate building in this study. The input and target features of these probabilistic PCE surrogates are summarised in Figure 2.…”

Multi-Horizon Data-Driven Wind Power Forecast: From Nowcast to 2 Days-Ahead

“…Particularly, Recurrent Neural Network (RNN) are Fig. 2: Neuron Structure of an LSTM, reproduced from [15] a subset of NN able to establish temporal sequences; making them best suited for time-series forecasting. Lastly, LSTM, are a subset of RNN, whose particularity is to reduce the incidence of the vanishing gradient problem.…”