Introduction
In complex decision-making environments such as finance, energy markets, and robotics, optimal actions today depend on a range of possible future trajectories. Traditional forecasting models, including Vector Autoregression (VAR) or Long Short-Term Memory (LSTM) networks, often struggle to capture the multimodal nature of future uncertainties. While modern diffusion-based generative models excel at producing high-fidelity probabilistic forecasts, they typically generate an unstructured set of scenarios. This paper introduces the Diffusion Scenario Tree (DST) framework, which bridges the gap between high-fidelity diffusion forecasts and the structured requirements of multistage stochastic optimization.
The DST Framework
To utilize probabilistic forecasts in Stochastic Model Predictive Control (MPC), information must be organized into a Scenario Tree. This structure is critical for enforcing "non-anticipativity"—the logical constraint that decisions made at a specific time cannot rely on information revealed in the future.
The authors propose a recursive algorithm to construct these trees:
- Probabilistic Forecasting: An autoregressive diffusion model (TimeGrad) samples multiple future trajectories conditioned on historical data.
- Clustering for Structure: At each stage, sampled trajectories are grouped via K-means clustering. The resulting centroids serve as the tree's nodes, representing representative future states.
- Probability Assignment: Branch probabilities are determined by the density of samples within each cluster.
- Pruning: To maintain computational tractability, only the top-$L$ child nodes with the highest cumulative probability are retained.
- MPC Integration: The resulting tree is fed into a Multistage Stochastic MPC solver, which optimizes actions across all branches simultaneously.
Experimental Evaluation: Energy Arbitrage
The framework was evaluated on a real-world energy arbitrage task using 10 months of hourly electricity price data from the New York State day-ahead market. The objective was to manage a simulated Battery Energy Storage System (BESS), buying electricity when prices are low and selling when they are high, while accounting for battery degradation.
Results and Discussion
The DST-guided MPC was compared against several benchmarks, including Oracle MPC (perfect future knowledge) and traditional tree-generation methods based on LSTMs and VAR models.
As shown in the results table, DST-MPC achieved a total reward of 858.09, significantly outperforming LSTM-based trees (727.32) and VAR-based trees (684.58). Notably, DST-MPC performed remarkably close to the Oracle MPC (889.30), demonstrating its ability to capture complex market dynamics.
Key insights from the study include:
- Structure Matters: Organizing predictions into a tree yields superior policies compared to unstructured Monte Carlo sampling, even when using the same underlying diffusion model.
- Data Efficiency: DST produces more stable and effective policies than model-free Reinforcement Learning (RL) baselines, such as DQN, which struggled in this low-data regime (averaging only 26.92).
- Interpretability: Unlike "black-box" predictors, the scenario tree allows human operators to visualize the specific future scenarios the controller is hedging against.
Conclusion and Future Work
While the DST framework offers significant performance gains, the iterative denoising process of diffusion models introduces computational latency. Future research will explore faster sampling variants like DDIM or DPM-Solver++ to enable real-time applications. Additionally, the authors suggest moving toward "problem-driven" tree generation, where the tree structure is optimized specifically to satisfy the physical constraints of the underlying system.
Summary updated