The transition towards sustainable structural power sources motivates vibration energy harvesters that avoid both lead-based piezoceramics and petroleum-derived substrates. This work presents an interpretable machine-learning framework for a resonant harvester formed from a natural-fibre (flax/epoxy) biocomposite host plate instrumented with a lead-free piezoelectric patch, benchmarked against conventional PZT-5A. A coupled first-order shear-deformation finite-element model of the harvester is developed and validated against analytical and published solutions spanning static deflection, free vibration, piezoelectric actuation and the electromechanically coupled power response. A 1500-sample Latin-hypercube dataset over eight geometric, material and circuit variables and three transducers (PZT-5A, BaTiO3, BZT-BCT) trains four surrogate models; a Gaussian-process regressor is the most accurate (test R2 = 0.95) for the predicted power density. These results are model-based predictions that await experimental confirmation. Shapley additive (SHAP) analysis reveals that proximity to resonance and the load resistance dominate the response, whereas the material contribution is governed by the transduction figure of merit d312/ϵ33T rather than the piezoelectric strain coefficient alone. Notably, BZT-BCT underperforms PZT-5A despite its larger |d31|, because of its high permittivity, giving the ranking PZT-5A > BZT-BCT > BaTiO3 consistently across the figures of merit, the simulated power density and the data-driven importances. The framework provides a transferable, physics-consistent route to sustainability-aware co-design of biocomposite vibration energy harvesters.