Data-Driven Stochastic Optimal Control for Intraday Electricity Trading by Renewable Producers

Battle Count: 60

Highly relevant to quantitative trading in electricity markets. The paper provides a rigorous continuous-time stochastic control framework for optimal execution in intraday markets, with realistic modeling of market microstructure (gate closure, delivery windows, imbalance settlement). The jump-diffusion price model captures heavy-tailed behavior observed in electricity markets. The numerical methodology (IMEX scheme, operator splitting, monotone discretization) is directly applicable to other optimal execution problems. The benchmark comparison methodology (TWAP, perfect foresight) is standard in algorithmic trading evaluation. However, the application is specific to electricity markets and renewable producers rather than general financial instruments.

Very high implementation complexity. Requires solving a 4D nonlinear HJB PIDE with jumps, involving: (1) three-stage backward PDE system with different state dimensions per stage, (2) Lie operator splitting with IMEX time discretization, (3) semi-implicit Rosenbrock linearization with Picard iterations for the nonlinear Hamiltonian, (4) positive-type generalized finite differences for the mixed derivative with Jacobi diffusion, (5) differential formulation for the double-exponential jump operator with recurrence relations, (6) domain truncation with probabilistic bounds, (7) careful boundary treatment at production boundaries (x=0,1) and price boundaries. The monotonicity conditions (CFL-type constraints on grid ratios) require careful parameter selection. Multiple proofs establish well-posedness and convergence.

The paper provides detailed model specifications, parameter tables (Table 5.1), grid discretization details, boundary conditions, and complete proofs in appendices. Data is publicly available from SMARD.de. Code availability is stated as 'upon publication' via GitHub. The methodology is fully specified with explicit equations, numerical schemes, and parameter values. However, the price forecast construction uses a moving-average proxy rather than a true ex-ante forecast, and the evaluation is ex-post rather than fully out-of-sample.