Wind Power Forecasting techniques: ANN vs. ANN-CFD hybrid approach

Introduction

Wind power forecast is crucial in electrical grid balance. Nowadays, wind power forecasting is widely done using statistical tools. Those convert wind speeds produced by Numerical Weather Prediction (NWP) models to power production.
Statistical tools find connections between historical NWP output and observed power and apply them to forecasted NWP output to predict the power production. Observed power production data need to be cleaned and, the statistical tool recognizes relationships between NWP output and observed power production. Therefore, improper cleaned data can lead to high errors.
One type of statistical methods are Artificial Neural Network (ANN) which connect NWP model data as input to wind farm power production as output [1]. A different approach is set using Computational Fluid Dynamics (CFD) together with ANN. CFD simulation of the wind flow near the ground can describes the local wind field around the wind farm. It is a deterministic tool which dynamically describes the wind speed from the mesoscale level onto the wind farm level.
We present the case of a wind farm where both approaches are used. The two results are compared and the added value of each of the two approaches is highlighted.

Approach

Wind power forecast is performed using two approaches:
1. A single ANN applied to the NWP model output to calculate the power production of the whole wind farm or of a single turbine. It can be seen as an Artificial Neural Network Power Curve (ANN wind-power).
2. An ANN connects the wind of the NWP model to observed wind conditions on site and this corrected values are used as input to the CFD. This is what we call the “ANN wind-wind + CFD” method [2]. The nominal power curve is used to perform the power calculations.
It can be useful to notice that in the ANN wind-wind + CFD case, the ANN acts as in MCP (Measure Correlate Predict) where a correlation is created between NWP data and met-mast data.

Main body of abstract

The test case wind farm is sited in Southern Italy. It is composed of 24 Vestas turbines, V52 and V42. The terrain is extremely complex: the presence of mountains is important in all the directions and there are severe slopes. The highest point in the domain of the wind farm lies at about 1000 meters a.s.l., while the lowest one lies at around 400 meters.
The CFD calculation is performed solving Reynolds Averaged Navier-Stokes (RANS) equations with RNG k-epsilon turbulence closure on a domain describing terrain and roughness.
The NWP model is the Weather Research and Forecast (WRF) model, which uses 3 domains with increasing resolution up to 1 km. Time series are extracted from WRF's forecasts at the reference points mentioned above.
The ANN's are trained with cleaned SCADA measurements from the turbines. Half of the data available are employed for training, half for validation. Each turbine data set is cleaned on the requirement that the turbine itself is in production.
The performance is calculated using standard wind power errors calculations: Bias, R2 Correlation coefficient, NAME and NRMSE; and a deeper analysis of the two approaches behave at the time series level is presented.


Conclusion

The studied wind farm is sited in a highly complex area and two approaches of wind power forecast are presented and the performance of each is calculated and investigated.
The characteristics of the two approaches are highlighted providing insight about the pros and cons of the two approaches.
Special attention is given to the added value of deterministic approaches like CFD and how the description of the physical behaviour of wind can improve the wind power forecast.



Learning objectives
The audience will learn how ANN can be used for power forecasting together with CFD simulations or not.
The audience will learn about the performances and differences between statistical and deterministic tools in wind power forecasting.
The audience will learn how to perform wind power forecasting based on CFD.