Neural Network-Based wind speed Prediction and Evaluation Using the NREL Wind Integration National Dataset

Abstract

Accurate wide speed (WS) forecasting is an important factor in optimizing wind energy production andkeeping power in a stable condition with lower operation costs. Therefore, this study is going to comparedifferent machine learning (ML) models such as Support Vector Regression (SVR), Decision Trees (DT),Gradient Boosting Machines (GBM), Multi-Layer Perceptron (MLP), and ensemble techniques usingVoting and Stacking. These models are trained and tested by taking data from the National RenewableEnergy Laboratory's Wind Integration National Dataset, also known as the WIND Toolkit, which is ahigh-resolution turbine location dataset of wind resources for nearly all locations within the contiguousUnited States and significant portions of Canada and Mexico. The key performance metrics include testingusing MSE, RMSE, and R² for each model. In fact, the results confirm that the relative performance ofensemble methods has greater stability and forecasting accuracy, with the Voting model outperformingall the other individual models, thus capturing complex patterns in the WS data. Consequently, SHAPanalysis underlined capacity factor and longitude as geographic location features among the mostimpactful, entailing the importance of such variables in the prediction tasks related to wind. These resultsunderline the potentiality of ensemble techniques in view of robust estimation of WS, which would helpthe management of renewable energy in a more reliable way. Further, the work will focus on some deeplearning (DL) approaches and extra meteorological data in order to improve model performance for thesustainable integration of wind energy into the power grid.