Ghent University's S3TS Beats Standard Grid Planning by 51% on Nonlinear Problems
S3TS merges Monte Carlo Tree Search with stochastic scenario trees, letting energy planners handle nonlinear models and renewable uncertainty at the same time — something neither approach could do alone.
Takeaways
- New algorithm: S3TS extends Monte Carlo Tree Search with pre-built scenario trees, enabling planning under uncertainty with nonlinear models.
- Core gap filled: Existing methods handle either nonlinearity (MCTS) or uncertainty (stochastic MPC) — S3TS handles both simultaneously.
- Linear results: Achieves costs within 14% of the MILP-optimal stochastic solution on a tractable Belgian grid imbalance benchmark.
- Nonlinear results: Cuts costs by up to 51% vs. a rule-based baseline and 5.4% vs. deterministic MCTS, with the gap widening at higher compute budgets.
- Plug-in compatible: Uses the same scenario trees as stochastic MPC, so existing probabilistic forecasting pipelines require no modification.
- Open source: Code available at github.com/loud-mime/S3TS; extensions to CVaR and broader domains planned.
Energy grid operators face a planning problem that sits at an uncomfortable intersection: the models needed to accurately simulate modern power systems are increasingly nonlinear, while the uncertainty introduced by renewable energy sources demands explicit probabilistic reasoning. Until now, no single planning framework handled both well. A new algorithm from researchers at Ghent University, Beebop.ai, and Imperial College London aims to close that gap.
The paper introduces S3TS (Stochastic Scenario-Structured Tree Search), an extension of Monte Carlo Tree Search (MCTS) that replaces the algorithm's usual assumption of deterministic transitions with a pre-built scenario tree encoding probabilistic futures. The result is a planner that can reason over uncertainty without needing a closed-form mathematical model of the system it controls.
The planning gap that motivated this work
The energy sector has two dominant planning paradigms, and each has a blind spot. Mathematical optimization (think Mixed-Integer Linear Programming, or MILP) handles uncertainty well through multi-stage stochastic formulations, but breaks down when system dynamics are nonlinear , discontinuous price ladders, neural-network-based demand models, and mixed-integer constraints all push conventional solvers to their limits. Monte Carlo Tree Search, on the other hand, can handle arbitrarily complex simulators because it only needs to run a model, not differentiate through it. But standard MCTS assumes deterministic state transitions, making it fragile when future outcomes are genuinely uncertain.
While existing approaches can address either non-linearity (e.g., Monte Carlo Tree Search) or uncertainty (e.g., stochastic mathematical optimization), there is a lack of planning techniques capable of addressing both challenges simultaneously. S3TS is the proposed bridge.
What a scenario tree actually is
A scenario tree is a compact, branching representation of how an uncertain variable (like grid imbalance) might evolve over time. Instead of assuming a single future trajectory, it encodes a set of plausible futures, each with an associated probability. This is the standard input format for stochastic Model Predictive Control (MPC) , and crucially, S3TS reuses exactly the same scenario tree that a stochastic MPC would consume. That means any existing probabilistic forecasting pipeline can feed S3TS without modification.