Machine-intelligent co-optimization of load forecasting and EV scheduling using KAN and TA-V2GS to support VPP design

Abstract

Electric vehicles (EVs) are rapidly becoming an important part of the power system and pose a high level of uncertainty in the operation of the system while providing a largely untapped distributed energy storage opportunity. Current sequential forecast/schedule approaches can directly result in forecasting errors that impact EV scheduling, resulting in suboptimal grid performance and reduced Vehicle-to-Grid (V2G) penetration in a Virtual Power Plant (VPP) program. This paper introduces an integrated machine-intelligent co-optimization framework that integrates high-accuracy load forecasting and tariff-aware adaptive EV dispatch. The performance of five advanced forecasting models, namely, Long Short-Term Memory (LSTM), hybrid LSTM–XGBoost, TimeMixer++, Variational Mode Decomposition, and Kolmogorov–Arnold Network (KAN), is evaluated using the long-term load data of Andhra Pradesh (2017–2024). Under quasi-stationary conditions, KAN outperforms all compared models, achieving the best mean absolute error (158.2 MW), root mean squared error (209.7 MW), mean absolute percentage error (1.86%), and R2 (0.982), and has the highest anomaly resistance under the COVID-19 shocks. The KAN forecast is used to coordinate bidirectional charging/discharging of 90 000 heterogeneous EVs in a Tariff-Aware Adaptive Vehicle-to-Grid Scheduler that maintains strict State-of-Charge constraints. The framework delivers around a 17% reduction in peak demand, a 21%–25% reduction in daily costs, and an 8.6% Net Revenue Index through V2G arbitrage, resulting in cumulative monthly savings of approximately Rs. 50 crore. The optimality gap of only 17.8% and a stable V2G contribution of 17–18 GWh/month confirm near-optimal economic and operational performance for VPP-enabled power systems.