A cooling zone is a monitored region of a data center with its own thermal conditions, workload, and cooling controls. The operating decision is to supply enough cooling to protect equipment without wasting energy. Zones must be coordinated because airflow and heat can affect neighboring regions. section 9.7 section 9.0
In a simulated cooling system, each zone has a fan and a chilled-water valve. Their speeds and openings are continuous choices, so controlling several zones together produces a high-dimensional action space. A policy gradient approach can learn continuous actions directly rather than selecting from a discretized set. section 9.0 section 9.8
Eight connected zones
An example environment arranges eight zones in a ring, with two neighbors per zone. At each step, the controller observes eight features per zone: temperature and pressure deviations, workload, average neighbor temperature, outside temperature, entering water temperature, and the previous fan and valve settings. That gives 64 observed values and two controls per zone. section 9.7
Listing 9.2 Environment initialization and zone adjacency
class DataCenterEnv:
def __init__(self, config: Config,
batch_size: Optional[int] = None):
self.cfg = config
self.batch_size =
[CA]batch_size if batch_size is not None
else config.batch_size #A
self.device = torch.device(config.device)
self.adjacency = self._create_adjacency_matrix() #B
def _create_adjacency_matrix(self) -> torch.Tensor:
n = self.cfg.n_zones
adj = torch.zeros(n, n, device=self.device)
for i in range(n):
adj[i, (i - 1) % n] = 1.0 #C
adj[i, (i + 1) % n] = 1.0 #C
adj = adj / adj.sum(dim=1, keepdim=True).clamp(min=1.0) #D
return adj
The adjacency matrix makes the ring connection explicit: each zone is linked to its two neighbors, whose temperatures can contribute to its observation. section 9.7
The simulated step limits how quickly fan and valve settings can change and keeps them between zero and one. Its reward penalizes fan and valve energy use as well as temperature and pressure violations. An attention-based REINFORCE policy uses the zone observations to coordinate cooling decisions across the floor. section 9.7
