A hybrid wavelet transformer LightGBM model optimized by optuna–TPE for global irradiance forecasting in solar electric vehicle energy management

• A hybrid Wavelet–Transformer–LightGBM model is proposed for irradiance forecasting. • Optuna–TPE is employed for efficient hyperparameter optimization. • The model improves short-term global solar irradiance prediction accuracy. • The framework supports energy management in solar electric vehicle systems. Accurate short-term global irradiance (GI) forecasting is a key enabler for efficient photovoltaic (PV) energy integration, intelligent charging strategies, and real-time energy management in Solar Electric Vehicles (SEVs). However, the inherently non-stationary and stochastic nature of solar irradiance, driven by rapid meteorological variations and seasonal effects, makes reliable prediction particularly challenging. To address these limitations, this paper proposes a novel hybrid forecasting framework, termed WTX–TPE–LGBM, which integrates Discrete Wavelet Transform (DWT)–based multi-scale decomposition, an Encoder–Decoder Transformer for long-range temporal modeling, and a Light Gradient Boosting Machine (LightGBM) for residual error refinement. The DWT decomposes irradiance signals into approximation and detail components, effectively capturing both slow-varying seasonal trends and high-frequency fluctuations induced by cloud dynamics. The Transformer module exploits multi-head attention to learn complex temporal dependencies across multiple time scales, while the LightGBM stage enhances generalization by correcting nonlinear residual patterns. To ensure optimal performance and computational efficiency, all model components are jointly tuned using Optuna’s Tree-structured Parzen Estimator (TPE), a probabilistic Bayesian optimization strategy that offers faster convergence and higher sample efficiency than classical metaheuristic optimizers. Extensive experiments on long-term solar irradiance datasets from multiple climatic regions demonstrate that the proposed framework consistently outperforms state-of-the-art models, including LSTM, BiLSTM, CNN–LSTM, Transformer, and LightGBM. The hybrid integration of wavelet-based decomposition, attention-based temporal modeling, and gradient-boosted residual refinement achieves accurate forecasts under diverse atmospheric conditions, with MAE < 0.02, RMSE < 0.03, MAPE ≈ 1.9%, and R 2 > 0.985. Moreover, the framework shows robustness to irradiance non-stationarity, efficient TPE-based convergence, and low inference latency for real-time SEV management.