Comparing Regression Models with Enhanced Interpretability for Wind Speed Forecasting in La Guajira-Colombia

La Guajira-Colombia offers immense potential for wind and solar-based power generation, making it an ideal location for implementing sustainable energy solutions. However, accurately forecasting wind speed is crucial to support renewable power monitoring. This issue requires coding non-stationary and noisy time series within short-term scenarios and maintaining a proper interpretability stage to highlight relevant input patterns to support power system planning tasks. This paper presents an interpretable artificial intelligence framework for wind speed forecasting, focusing on data from La Guajira-Colombia. The analysis covers short-term horizons with hourly resolution. Besides, both classical Machine Learning and Deep Learning models are tested within a relevant analysis scheme based on L2-Norm and Layer-wise Relevance Propagation, allowing for a thorough evaluation of the efficacy of each method. Namely, k-Nearest Neighbors, Kernel Ridge Regression, Simple Recurrent Neural Networks, Gated Recurrent Units, and Long and Short-term Memory Networks are studied. Achieved results demonstrate the effectiveness of tested techniques within the initial prediction horizons. Still, the analysis also reveals an increase in error as the prediction horizon extends. Also, relevance analysis allows interpreting pertinent information regarding salient input samples to code non-stationary and noisy patterns. The code is available at https://github.com/ealeongomez/Comparing-Regression-Models-with-enhanced-Interpretability .