Posters

Abstract

The study targets the enhancement of accuracy and availability of 10-minute wind speeds for Continuous Wave Lidar Systems by leveraging machine learning techniques. Conventionally, 10-minute wind speeds are derived from averaging 1-second wind measurements, susceptible to errors due to various factors. Traditional filtering methods, often reliant on rigid thresholds, struggle to discern valid from erroneous data points, resulting in both false positives and negatives. The proposed solution introduces a machine learning-based filtering approach, employing a Bidirectional LSTM model. Trained on diverse data sets from multiple regions, the model discerns intricate patterns and dependencies, considering both preceding and succeeding elements. The sigmoid activation layer assigns a confidence value to each wind speed measurement, indicating its usability probability. Testing conducted at the Georgsfeld site in Germany, spatially independent from the training data, reveals promising results. The model enhances availability and accuracy across all heights, showcasing potential enhancement in slope, R2, RMSE, and RMSD. The observed performance suggests the generalisability of the filter to new, unseen locations. However, acknowledging the study's limitation in evaluation at a single test site, broader claims regarding the filter's usability are cautioned.