CUP-EV: Hybrid DLinear-based cost and uncertainty-aware probabilistic multi-horizon electric vehicle charging forecasting

Accurate and uncertainty-aware forecasting of electric vehicle (EV) charging demand is critical for grid stability, operational planning, and cost-efficient coordination of charging infrastructure. However, existing studies often emphasize deterministic forecasts, limited horizons, or computationally intensive architectures without explicitly incorporating cost-sensitive objectives or rigorously evaluating probabilistic calibration, which limits real-world applicability. This paper addresses multi-horizon probabilistic EV charging demand forecasting under uncertainty and cost-aware evaluation criteria. We propose the Cost and Uncertainty-aware Probabilistic EV forecasting (CUP-EV) framework, built upon a lightweight hybrid architecture, DLinear-Res-Prob, which integrates linear trend decomposition with a nonlinear residual multilayer perceptron (MLP) through a learnable adaptive mixing mechanism to generate quantile forecasts for 1, 3, 6, and 24 hours ahead. Reliability diagnostics, including coverage-nominal analysis and Probability Integral Transform (PIT) evaluation, are conducted to assess calibration, and split conformal adjustment is applied to correct dataset-dependent under-coverage without retraining. Experiments on the Jiaxing and Palo Alto datasets demonstrate that the proposed model outperforms eight competitive baselines, achieving a reduction of up to 9.76% MAE and 17.86% asymmetric cost improvement in Jiaxing, while maintaining consistent RMSE and cost gains in Palo Alto. The model uses fewer than 0.3M parameters, enables sub-millisecond inference, and trains substantially faster than transformer-based architectures.