Rethinking Efficiency: The Hidden Cost of SAC in HVAC Control

In a recent exploration of Soft Actor-Critic (SAC) applications in HVAC control, researchers have pinpointed a essential inefficiency. The energy floor, or the lowest cost achievable given action space constraints, stands at a notable USD 35.51 per day. This cost is nearly entirely attributed to continuous electrical loads, accounting for USD 35.44 of the total, with gas consumption barely making a dent.

The Challenge of Baselines

Interestingly, the standard SAC baseline, when initialized with schedule-policy replay buffer transitions, converges at a higher cost of USD 37.18 per day. That's 4.7% above the energy floor. This gap highlights a significant inefficiency in the current baseline approach. The primary culprit? Buffer initialization, which stands as the dominant source of sub-optimality in this setting. When starting from an empty buffer, costs drop dramatically to USD 35.57 per day, effectively closing 96% of the gap.

Constraints and Opportunities

One might think increasing the supply water temperature range could yield savings. However, expanding it by 10 K only offers a minuscule USD 0.03 daily saving. Any further expansion risks breaching physical constraints. This reveals an inherent limitation in equipment efficiency that algorithms alone can't overcome. The real question is, how much can we optimize before hitting these brick walls?

Discount Factor Dilemma

Another fascinating finding points to a discount factor coupling, with an effective gamma of 0.891, reducing the planning horizon from 8.3 hours to a mere 46 minutes. This issue isn't isolated to one setup but is a benchmark-wide problem that demands an audit. Should the industry re-evaluate these standards to unlock potential savings and efficiency gains?

The Bigger Picture

A comprehensive ablation study confirmed that configurations with a pre-filled buffer cluster tightly within a narrow range (USD 37.18 to USD 37.42). The takeaway? Equipment minimum power, not algorithmic design, imposes the binding constraint. While SAC shows promise, the real limitation lies within the hardware.

What they did, why it matters, what's missing. This research underscores the necessity for a dual approach to optimization: refining algorithms while acknowledging and addressing physical constraints. The question remains, how do we move beyond these barriers to achieve true efficiency in HVAC systems?