Wind Power Forecasting Based on Deep Neural Networks and Transfer Learning

State-of-the-art forecasting systems are often based on single machine learning models that have been trained for individual wind farms. The data of a each wind farm is typically used exclusively for its ""own"" model. This article presents two approaches with deep neural networks to make the data usable across wind farms with transfer learning. In the first case, the adaptation to individual wind farms is achieved by an separate output layer for each wind farm. With the second approach, a Bayesian wind farm embedding is proposed. An experiment with realistic forecast conditions based on power measurements and weather forecasts of 19 wind farms is carried out. The proposed techniques are compared to established single wind farm models such as random forests, gradient boosted regression trees, and simple multi layer perceptrons. Our results indicate that a significant improvement in prediction quality can be achieved using multi-task learning, especially with a short time span of historical training data.