Application of Machine Learning (Knn) With Autohyperparameter Tuning for Hydrogen Production Probabilistic Forecasting From Wind Energy
Abstract
One of the main challenges for the energy transition is knowing how to deal with the variability of solar and wind sources, as this poses risks for renewable engineering projects. In this sense, this work presents the application of self-adjusting tuning for the k-Nearest Neighbors (kNN) machine learning model in a wind energy and green hydrogen production scenario. The methodology adopted consisted of a literature review to develop time series ensembles, and the results obtained were compared with the literature on the same subject. Three types of analyses were performed: wind speed prediction, direct prediction of green hydrogen production, and indirect prediction of hydrogen using the optimized values of the lag columns and kNN hyperparameter for wind speed. The distinctive feature of this approach is the combination of the bagging mechanism with the algorithm developed to auto-adjust the hyperparameter “n_neighbors” until the model stabilizes without supervision, which is proved useful and applicable to Artificial Intelligence models with more hyperparameters, allowing simulations to be performed for a large number of models and lag columns (values from previous time steps). The proposed methodology obtained better results in terms of prediction error (RMSE and nRMSE) and optimal values for R², indicating advances over existing approaches compared in the literature.
Keywords
wind energy, green hydrogen, machine learning, self-adjusting tuning, hyperparameter optimization