An adaptive LASSO-GARCH framework for robust EV parking occupancy forecasting and uncertainty management

This paper is dedicated to predicting the occupancy capacity of public parking lots for electric vehicles (EVs). The proposed method is a hybrid framework based on adaptive regression and dynamic volatility modeling. This method begins with extracting neighbors based on similarity to reduce the input dimensions. Then, the expected point load is estimated using least absolute shrinkage and selection operator (LASSO) regression. The main innovation of this method is the design of a two-stage mechanism that, in the early hours of the day due to data scarcity (cold start phase), utilizes information from similar days. Then, as the system stabilizes, the model smoothly connects to the generalized auto-regressive conditional heteroskedasticity (GARCH) algorithm to track the day's continued volatility. Additionally, a physical processing step has been subsequently added to yield correct, discrete, and non-negative outputs. The proposed framework has been validated using a comprehensive dataset based on real data from Caltech ACN-Data. Comparative analyses demonstrate the superiority of the model over benchmarks such as ARIMA and advanced deep learning architectures (LSTM and GRU) in terms of relative accuracy and stability (RMSE equal to 2.71 EVs). The analysis shows that the model has an absolute error of zero in 50% of cases. Additionally, the estimation error remains below 6 EV during 95% of working hours (while the total daily traffic of the parking is 400 EVs). Additionally, a degradation study has been presented to illustrate the necessity of synergy between the components of LASSO-GARCH. The results show that this integration has led to point accuracy. It also optimizes the uncertainty bounds by reducing the mean prediction interval width (MPIW) and maintaining a 95.5% prediction interval coverage probability (PICP).