MPC-EP replay predictor may diverge from the main EnergyPlus trajectory

[xudongwu-0]

Problem

The current mpc_ep baseline uses EnergyPlus as a short-horizon replay predictor, but it does not clone the full internal state of the main running EnergyPlus simulation.

Each candidate rollout starts a fresh EnergyPlus run, goes through warmup, and replays from Day 1 to the current MPC decision time. Although the replayed objective states are now aligned to the live decision-time temperature and meter power, the future H-step trajectory can still diverge from the main simulation.

Evidence

Recent diagnostics show that the live/replay decision-time temperature and first-step power are already very close, but the predicted H-step facility power can still differ from the realized main-run meter trace.

Example diagnostic result:

• role_a_mpc_ep_H6_tianjin_3days
• MAE around 0.8 kW
• RMSE around 1.24 kW

This suggests that the remaining error is likely from future HVAC/device cycling divergence, not simply from missing current zone-temperature inheritance.

Why This Matters

This makes mpc_ep less like a true EnergyPlus oracle and more like an EnergyPlus replay-based predictor. As a result, mpc_dynamic can sometimes perform better or more stably, even though mpc_ep seems physically richer.

Possible Directions

• Keep documenting mpc_ep as an EnergyPlus replay-based horizon predictor, not a full-state oracle.
• Investigate whether EnergyPlus internal state cloning or warm-starting is possible.
• Add additional state synchronization for devices if available.
• Compare predicted vs. realized H-step trajectories routinely in benchmark reports.
• Track whether prediction divergence is dominated by HVAC cycling, water-heater state, EV state, or other controllable appliances.