C2SLM: A Correlation-based Clustering-assisted Sparse Learning Model for Electric Vehicle Market Demand Forecasting

Abstract

Battery electric vehicle (BEV) market demand forecasting is challenged by high-dimensional feature selection from a regression model. The Least Absolute Shrinkage and Selection Operator (Lasso) is a sparse learning model used for this purpose. However, due to the correlation structure of data, Lasso can become unstable in its feature selection. This instability is further amplified in econometric panel data because of unobserved heterogeneity. Motivated by this challenge, we propose a correlation-based clustering-assisted sparse learning model (C2SLM) based on Lasso for a BEV market demand forecasting problem in the U.S. state of Alabama. Three stages structure the C2SLM: (i) correlation-polarized feature generation, (ii) panel-based feature clustering, and (iii) clustering-assisted Lasso (caLasso). We specifically aggregate the data to capture heterogeneous market factors for the panel. We test the C2SLM against benchmark models with or without panel consideration. Our proposed model outperforms others and sheds significant insights into BEV policymaking. Additionally, our model performs consistently robustly on Gross Domestic Product (GDP) and S&P 500 stock trading market datasets.