A handheld manipulator offers a feasible way to relieve the challenge of continuous curvilinear capsulorhexis (CCC) by controlling peeling force actively. The precise identification and prediction of voluntary peeling force (0-2 Hz) form the basis of force controller design. However, achieving this goal is limited by species differences (experimental validation: animal model and clinical surgery: in vivo human). This article proposes a framework to relieve this problem. The proposed framework consists of a wavelet-based process module and a prediction module based on long short-term memory (LSTM) and the Transformer's encoder. Peeling forces measured from ex-vivo pig eyes (no treatment, treated with 0.1% and 0.5% sodium hypochlorite), in vivo rabbit eyes, and ex-vivo human lenses are utilized to validate the proposed framework. Experimental results show that the root-mean-square (rms) error of identification is 0.37 mN. The rms error of prediction is 0.41 mN (steady) and 0.87 mN (sharp, prediction time: 50 ms), respectively. The rms prediction error of ex-vivo human samples is 0.36 mN (steady) and 1.05 mN (sharp), even though all training data come from animal models.