Bayesian neural network-based policy effect prediction for green transformation of power business environment

Predicting how green policies reshape power business environments remains notoriously difficult. The underlying dynamics are nonlinear, the uncertainties substantial, and conventional models often fall short. This study develops a Bayesian neural network framework designed specifically for forecasting and optimizing green policy outcomes within the Fujian power system, placing particular weight on quantifying prediction uncertainty to support sound decision-making. Our methodology weaves together stochastic variational inference and multi-objective optimization, thereby capturing the channels through which policies transmit their effects to environmental outcomes. Drawing on empirical data spanning 2018–2024, we find that this approach outperforms standard machine learning techniques by roughly 4–5% points in prediction accuracy while delivering markedly better uncertainty calibration. Scenario analyses reveal that moderate-to-high policy intensity tends to achieve favorable cost-effectiveness, with renewable energy incentives, carbon pricing, and regulatory enforcement standing out as especially potent drivers of transformation. Perhaps more importantly for practitioners, the framework demonstrates that well-designed moderate-intensity strategies can surpass maximum-intensity approaches once diminishing returns enter the picture. By enabling joint assessment of environmental gains, economic efficiency, and operational stability under uncertainty, this work offers a practical foundation for evidence-based policy design—though readers should bear in mind that our validation remains grounded in the Fujian regional context.