Abstract: The GM(1, 1) power model is frequently employed for forecasting nonlinear data, yet its parameter estimation faces challenges of slow convergence and susceptibility to local optima in existing methods. This study proposes an enhanced GM(1, 1) power model optimized via stochastic gradient descent (SGD), incorporating two innovations: dynamic background-value weights replace fixed ones to adaptively capture data trends, and full-information optimized initial conditions mitigate errors from insufficient information utilization. The SGD algorithm simultaneously optimizes the power exponent, background-value weights, full-information initial condition weights, and model parameters. Comparative experiments with particle swarm optimization (PSO) applied to the same model reveal that SGD, despite marginally slower convergence, achieves significantly higher precision and exhibits superior capability to escape local optima.