The load of the elastic tooth drum-type pepper harvester is a key parameter affecting harvesting efficiency and quality. Real-time analysis and prediction of drum load are crucial for stabilizing harvester operation and optimizing performance. Existing research focuses on either machine vision-based image analysis, which is difficult to collect in the field, or parameter-mapping methods, which suffer from time lag. This study proposes a GARCH-KPCA-ATLSTM method for load prediction, combining the generalized autoregressive conditional heteroskedasticity (GARCH) model, kernel principal component analysis (KPCA), and attention-enhanced long short-term memory (ATLSTM). EMD is first applied to denoise and reconstruct the load signal, removing mechanical vibration and other interferences. Conditional heteroskedasticity is confirmed, and the GARCH series (one symmetric and three asymmetric models) is introduced to extract fluctuation features. KPCA reduces dimensionality, removing redundant information and saving 2.91 s in computation while slightly improving accuracy. Additive attention in LSTM emphasizes critical information, enhancing learning of nonlinear relationships and further improving prediction. Comparative experiments demonstrate the model’s reliability. The method achieves RMSE = 0.911, MAE = 0.682, MBE = −0.025, MAPE = 1.147%, R² = 0.968, with a runtime of 2.023 s, confirming high accuracy and stability. This study provides a theoretical and technical foundation for real-time load prediction of pepper harvesters.