The grey power model captures nonlinear trends, but its accuracy heavily depends on the power exponent, which is often empirically selected. For instance, the grey Verhulst model and GM(1,1), two special  cases of grey power models and widely applied,  struggle to adapt to the fitting of complex data. To address this, we propose an Improved Grey Power Model (IGPM) that optimizes the exponent γ via a dual-path optimization strategy integrating grid search and the Artificial Fish Swarm Algorithm (AFSA) within the theoretically determined bounds of [-2, 2]. In the context of small-sample data, validated on annual power generation and metal fatigue strength, the optimized IGPM reduces MAPE by 37.5% and RMSE by 38.0% on average over the standard GM(1,1). The framework provides a robust solution for data-scarce engineering predictions.