Data Driven Modeling and Prediction of the Power Generation of a Wind Farm by Mamba Learning

With the rapid development of wind energy, accurate prediction of the output electric power of wind farms has become the key to energy management for grid stability. Although the technologies in harvesting wind power have advanced significantly in various aspects, there are still many challenges hindering the further improvement of the power generation efficiency, among which wake effect is a major one. To overcome this challenge, models based on first principles, e.g., fluid dynamics, are built in the literature, which are too complicated to be used in real time, and still suffer from significant inaccuracy. This has motivated the development of AI based models to describe the dynamic processes of wind farms. Despite the avoidance of first-principle modeling, the existing AI models still do not possess sufficient accuracy and efficiency to deal with the complex wake phenomenon. Recently, a new technique, namely Mamba, has been developed for modeling long sequence voice data with comparable performance to Transformers, but only with linear complexity. Due to these advantages, Mamba is especially suitable for modeling and estimating complex wake effects on a wind farm. This work is thus devoted to exploring its effectiveness in fulfilling this task. Moreover, to capture the abrupt temporal and spatial changes of the wake, a local attention mechanism is incorporated into the original Mamba model to improve its ability to capture local temporal dependencies and short-term fluctuations. The developed model is trained and tested with the data generated from FLORIS, a widely used simulation package developed by NREL. Comparisons with other AI models demonstrate the advantages of the developed Mamba model in predicting the power generation of a wind farm.