Improved Short-Term Wind Power Forecasts: Low-Latency Feedback Error Correction Using Ramp Prediction and Data From Nearby Farms

Abstract

This paper shows that the short-term wind power forecasts of a target farm can be significantly improved by using ramp predictions and information from nearby farms. To do this, we first obtain benchmark wind power forecasts from Scipher.Fx by Utopus Insights, Inc, which is owned by Vestas Wind Systems. Second, we build a low-latency feedback error correction model that predicts the forecast error at a given look-ahead time based on a novel ramp predictor, the last known forecast errors, and optionally, the last known forecast errors from nearby farms. The predicted forecast error is then combined with the benchmark wind power forecast to obtain the improved forecasts. The novel ramp predictor is constructed using the benchmark wind power forecast and optionally, measured data over a defined time window, to improve the less accurate wind power forecasts during ramp events. The ramp predictor also improves forecast accuracy for longer look-ahead times by a second mechanism which we detail. The nearby farm selection algorithm is based on two approaches: 1) Correlation analysis of historical data, and 2) Feature selection based on Shapley additive explanations feature importance values. Our approach was tested on 17 wind farms in Europe and the results showed that the ramp predictor can decrease the average relative normalized mean-absolute error of 10 minutes to 6 hours look-ahead forecasts by 3.61%. Additional improvements from nearby farms can be as high as 2.54% for some look-ahead times depending on the availability of data from upwind farms.

Keywords

• Wind power forecasting
• feedback error correction
• machine learning
• XGBoost
• wind power ramp prediction
• correlation analysis