CRISPR-DEO: Decision-Aware Economic Dispatch Optimization via Sparse Gradient Editing for Power System Forecasting

Power systems face a forecasting paradox: models optimized for statistical accuracy often produce suboptimal operational decisions. We address this misalignment through a decision-aware economic dispatch optimization framework that trains forecasting architectures using gradients derived from actual dispatch costs rather than isolated prediction errors. By integrating probabilistic forecasting with risk-constrained economic dispatch via differentiable optimization, the framework enables neural architectures to internalize systemic operational regret directly. Inspired by CRISPR gene editing, we introduce a sparse gradient modification method updating only 10% of network parameters per training iteration. Our approach, in contrast to traditional magnitude-based pruning, chooses parameters according to how sensitive they are to forecast entropy and downstream dispatch costs, guaranteeing that updates are in line with operational goals. Continuous relaxation maintains end-to-end differentiability with an optimality gap below 5%. Validation using a decade of hourly data from South Africa’s national grid demonstrates operational cost reductions of nearly 40% compared to deterministic baselines and over 14% versus decoupled approaches. The framework maintains superior forecast accuracy while achieving 90% parameter sparsity and improved risk stability. Theoretical analysis reveals negative generalization gaps in which test performance exceeds training performance through reduced model complexity. Enhanced integration of renewable energy delivers two benefits: yearly operating savings that exceed $11 million for grid operators and large reductions in carbon dioxide emissions. With inference latencies under 2 seconds on standard hardware, the framework provides a computationally feasible pathway for decision-aware economic dispatch optimization in power systems where asymmetric operational costs and tail-risk management are critical.