A Singular Control Problem for Data Center Electricity Cost Minimization

The paper provides a complete analysis of a stochastic control problem where the cost to minimize involves the running maximum of the underlying controlled state process. The model is motivated by the electricity cost minimization of a data center facing coincident peak charges and possibly benefiting from participation in a demand-response program.

The paper studies a single-agent continuous-time stochastic control problem over a one-dimensional controlled diffusion process with a lower reflecting boundary. This process, denoted (Lt)0≤t≤T, represents the electricity demand (load) at time t of a data center trying to minimize the cost of its electricity consumption over the period [0, T]. The coincident peak charges are modeled via an integral with respect to the running maximum Ms of the load Lt for t ≤ s. The underlying state of the optimization problem is (Ls, Ms), where Ls is the load process and Ms = supu≤s Lu its running maximum.

The cost functional includes three sources: a running cost function f(t, l, δ) = (a + bl − α0)δ + (β/2)δ2, a coincident peak penalty of the form κ∫ φ′(Ls) dMs, and a demand-response reward of the form ∫ Πs(Bs − Ls)+ ds. The value function is defined as v(t, l, m) = inf J(t, l, m; δ) with v(T, l, m) = 0.

The main mathematical challenge comes from the inclusion of the running maximum in the state of the system, forcing the original problem into a singular stochastic control problem on a closed wedge in the plane with reflection on parts of the boundary. The paper proves that the value function is locally Lipschitz continuous in space using a gap-monotonicity estimate for coupled reflected loads and a deterministic total-variation bound for the difference of two running maxima.

The paper derives the dynamic programming principle and viscosity formulation for ordinary progressively measurable controls. The corresponding HJB equation includes a reflecting Neumann condition on the lower boundary for the load direction as well as an oblique boundary condition on the diagonal part of the boundary of the domain due to the effect of the running-maximum process. The paper proves existence of a solution in the viscosity sense as well as a comparison theorem providing conditional uniqueness in the corresponding viscosity class.

The paper also provides numerical results illustrating the trade-off between the two conflicting incentives comprising the expected cost of the data center. Three computational routes are compared: incremental dynamic programming, deep Galerkin value approximation, and direct SGD optimization.

Improvements to AI Systems Based on This Paper:

1. AI for Real-Time Demand-Response Optimization in Data Centers

• Improvement: Train a reinforcement learning (RL) agent using the paper’s stochastic control framework, where the state includes both current load and its running maximum (to capture coincident peak penalties).

• Capability: The AI can dynamically adjust server workloads, battery storage, or deferrable tasks to minimize electricity costs under uncertain demand, while explicitly avoiding peak spikes that trigger high charges—something standard RL ignores.

2. AI for Peak-Shaving in Smart Grids with Memory-Dependent Costs

• Improvement: Incorporate the paper’s HJB equation with oblique boundary conditions into a neural-network-based solver (e.g., deep Galerkin method) to learn optimal control policies for energy storage or flexible loads.

• Capability: The AI can handle cost functions that depend on historical maxima (e.g., demand charges, grid congestion fees), enabling more accurate long-horizon planning than myopic or memoryless models.

3. AI for Robust Control Under Path-Dependent Constraints

• Improvement: Use the paper’s viscosity solution and comparison theorem to design AI controllers that are provably stable even when the state space has non-smooth boundaries (e.g., load limits, capacity caps).

• Capability: The AI can guarantee performance bounds in safety-critical applications (e.g., hospital power systems, manufacturing) where violating a peak threshold has severe penalties.

4. AI for Multi-Source Cost Minimization with Conflicting Incentives

• Improvement: Apply the paper’s numerical comparison (incremental DP vs. deep learning vs. SGD) to build an AI that automatically selects the best optimization algorithm based on problem size and nonlinearity.

• Capability: The AI can adaptively trade off between computational speed and accuracy for real-time decisions, balancing demand-response rewards against peak penalties.

5. AI for General Path-Dependent Stochastic Control

• Improvement: Extend the paper’s gap-monotonicity and total-variation bounds to train AI models that generalize across different stochastic processes (e.g., financial asset prices, network traffic) with running-maximum costs.

• Capability: The AI can solve a broader class of problems—like optimal execution with price impact or inventory management with stockout penalties—where the cost depends on the worst-case historical state, not just the current one.

6. AI for Data Center Demand-Response Bidding

• Improvement: Use the value function from the paper to train an AI that bids into demand-response programs (e.g., capacity markets) by predicting the optimal load reduction under uncertainty.

• Capability: The AI can decide how much load to shed or shift in advance, maximizing reward while avoiding coincident peak charges, even when future demand is stochastic.

7. AI for Explainable Peak-Risk Management

• Improvement: Leverage the paper’s explicit boundary conditions to build an AI that provides interpretable rules (e.g., “keep load below threshold X for the next 2 hours”) derived from the viscosity solution.

• Capability: The AI can offer human-readable operational guidelines for grid operators or facility managers, improving trust and adoption in real-world energy systems.