Creative Commons

Wind is quite a peculiar source of energy because of its availability, energy generating potential, and intermittency. Most of the problems associated with wind power stem from its intermittent nature including operational and control challenges that disrupt grid reliability [1], problems in contributing towards economic dispatch [2], and inability to participate in the reserve market [3]. Accurate forecasting of wind energy is an effective tool to tackle the intermittency problem, helping to determine net demand, power flows, voltage and frequency deviation. Wind Energy produced depends on weather conditions, such as temperature, wind speed, and humidity, and condition of the turbine and generator such as generator, gear and bearing rpm, fluid temperature, and particles in the oil. In recent years there has been to shift to machine learning methods for wind energy forecasting because of their ability to deal with complex problems. However, as more complex algorithms are being developed, their interpretability and explainability is decreasing. These models are called black boxes. This is not acceptable when dealing with time sensitive tasks or when any errors in prediction can prove fatal. Explainable AI (XAI) is a field that aims to open this black box by making models more explainable, acting as a bridge between machine and human understanding. Explainability helps in all stages of the machine learning pipeline and establishing trust between the user and system.The aim of this thesis is to compare the performance of Linear Regression, XG Boost, Random Forest and Neural Networks for prediction of wind power. Two different datasets are used to understand which variables are more effective for the prediction. Three popular XAI techniques, Feature Importance, LIME and SHAP, are used to understand the features that the models consider most important while making the prediction, both locally and globally. The findings from XAI techniques are incorporated into the models to try to understand if the variables identified as most influential are sufficient to make the predictions and how do the models perform if the most important variables are removed from the datasets.