Posters

Abstract

The high spatial and temporal variability of the wind and solar resources can result in additional costs for the power plant owners due to energy imbalance penalties imposed by the balancing authorities. To reduce these penalties, stakeholders must turn to additional resources such as energy storage systems or they must improve the accuracy of their power forecasts. This study focuses on improving short-term wind power forecasts. Our approach to improved short-term wind power forecasts uses feedback error correction. We first established benchmark wind power forecasts using a commercially available forecast application (Scipher.Fx). Second, we developed a machine learning-based bias correction algorithm to predict the forecast error for a given look-ahead by using the last observed forecast errors and a novel ramp predictor at the target farm, together with the forecast errors occurring in nearby farms. The predicted forecast error is then combined with the benchmark forecast to obtain the corrected wind power forecast. In the course of our research, we explored several approaches to constructing the ramp predictor using power measurements and power forecasts from our benchmark forecast application. We also explored methodologies for the selection of nearby farms to best account for the complex spatial-temporal patterns in error propagation. The proposed method was tested in 17 wind farms located in Northern Europe. The results showed an average 3.6% relative decrease in nMAE for the 10 minutes to 6 hours look-ahead forecasts over the whole test portfolio. A significant portion of this improvement came from implementing the ramp predictor, which only uses information available at the target farm. In fact, through the implementation of the ramp predictor, the results show a more significant average decrease in nMAE during ramp-up and ramp-down events, periods with typically higher imbalance penalties for stakeholders. The additional improvement came from the information from nearby farms. The geographic distribution of the farms in the test portfolio led to some target farms drawing more value from nearby farms' information than others. The findings suggest that nearby farms located upwind from the prevailing wind direction can improve forecast accuracy at the target farm. Furthermore, results for our test portfolio indicate that information from surrounding farms proved more valuable for shorter look-ahead times (from 10 minutes to 3.5 hours). Low-latency access to data from many geographically distributed sensors will be key to the practical implementation of this methodology.