Long-Term Carbon-Efficient Planning for Geographically Shiftable Resources: A Monte Carlo Tree Search Approach

TL;DR

The paper addresses reducing carbon emissions in power systems by strategically siting and operating geographically shiftable loads over long horizons. It develops a long-term planning model formulated as a large mixed-integer program that accounts for emissions from fuel generation, renewable curtailment, and load shifting, using fine-grained 20-year scenarios. To solve the resulting scale and combinatorial complexity, it introduces an adapted Monte Carlo Tree Search (MCTS) with an Iterative Priority Tree (IPT) representation, enabling parallelizable per-time-step subproblems and anytime stopping. Empirical results on networks up to 1888 buses show more than 10% carbon reduction with up to 8.1× speedups over standard solvers, demonstrating the approach’s scalability and practical impact for carbon-aware infrastructure planning.

Abstract

Global climate challenge is demanding urgent actions for decarbonization, while electric power systems take the major roles in clean energy transition. Due to the existence of spatially and temporally dispersed renewable energy resources and the uneven distribution of carbon emission intensity throughout the grid, it is worth investigating future load planning and demand management to offset those generations with higher carbon emission rates. Such techniques include inter-region utilization of geographically shiftable resources and stochastic renewable energy. For instance, data center is considered to be a major carbon emission producer in the future due to increasing information load, while it holds the capability of geographical load balancing. In this paper, we propose a novel planning and operation model minimizing the system-level carbon emissions via sitting and operating geographically shiftable resources. This model decides the optimal locations for shiftable resources expansion along with power dispatch schedule. To accommodate future system operation patterns and a wide range of operating conditions, we incorporate 20-year fine-grained load and renewables scenarios for grid simulations of realistic sizes (e.g., up to 1888 buses). To tackle the computational challenges coming from the combinatorial nature of such large-scale planning problem, we develop a customized Monte Carlo Tree Search (MCTS) method, which can find reasonable solutions satisfying solution time limits. Besides, MCTS enables flexible time window settings and offline solution adjustments. Extensive simulations validate that our planning model can reduce more than 10\% carbon emission across all setups. Compared to off-the-shelf optimization solvers such as Gurobi, our method achieves up to 8.1X acceleration while the solution gaps are less than 1.5\% in large-scale cases.

Figures (9)

• Figure 1: The schematic of our proposed spatial demand shifting framework. The shifting will be conducted among the selected locations to minimize the carbon emission of the whole system.
• Figure 2: Illustration of IEEE 14-bus system and load shifting across regions. Renewable generation such as wind generation, with a assumed carbon intensity of 0, is prioritized for minimizing system-level carbon emissions. Bus 5 has the highest carbon intensity of generation consumption, while buses 1 and 10 have lower carbon intensities or surplus wind generation, and direct power provision is constrained by the line limits. Thus, the controllable load of bus 5 is shifted to buses 1 and 10 to optimize carbon emissions.
• Figure 3: Illustrative example of MCTS-based location for a 3-bus system. We demonstrate the process of the 5th search round. (a) explains the 4 MCTS phase of one learning round. (b) show the specific sampling and reward calculation process in the simulation stage. (c) gives the tree structure of the IPT. Based on the given assumptions for $N_i$ and $V_i$, node G is determined to have the highest UCB value, and therefore, it is selected as the next node to visit. In this round, node G is visited but not expanded, resulting in the addition of its two child nodes, G1 and G2, to $\mathcal{T}_{MCTS}$. These child nodes are initialized with '$N_{G1}=0, V_{G1}=0, N_{G2}=0, V_{G2}=0$'. The action 'located at bus 1' is randomly chosen, leading to the selection of node G1 for the simulation stage. However, since there are no available actions at node G1, the simulation terminates at the state where the location decision is $\boldsymbol{z}$ = [1,1,0]. This indicates that bus 1 and bus 2 are selected for load shifting. \ref{['LS: MODEL_LP']} can be solved in parallel to reduce the solution time. The reward for node G1 in this round is assumed as $\gamma_k = 30$, which is set as the sum of optimal objectives of \ref{['LS: MODEL_LP']}. The UCB values for $\mathcal{T}_{MCTS}$ are updated through backpropagation: $N_{G1}=1, V_{G1}=30, N_{G}=2, V_{G}=50, N_{r}=105, V_{r}=5$. UCB values for other nodes remain the same. Given a sufficient time, $\mathcal{T}_{MCTS}$ can involve the most promising region of $\mathcal{T}_{IPT}$ and there will be a converged path start from the root to a leaf node, which can get a close-to-optimal objective.
• Figure 4: The average levels of the normal load, wind generation, and controllable load for per node in the investigated systems. The values represent the peak values within a day and the profiles are scaled to different levels according to the scales of systems.
• Figure 5: The hourly profiles within a day.

Theorems & Definitions (3)

• Proposition 1
• proof
• Remark 1