Learn to chill

Abstract

Centralized chiller plants with multiple chillers are typically over-provisioned. Therefore, intelligent scheduling is required for the supply (operating chillers) to efficiently meet the demand (actual cooling load of buildings). Traditional cooling-load based control (CLC) may result in poor part-loaded efficiency. Recent data-driven approaches to chiller control either unrealistically assume perfect knowledge of individual chiller power at various leaving chilled water temperatures (LWTs) or control all chillers with same LWT. We complement existing work with iChill, an end-to-end learning-based intelligent chiller power prediction and scheduling strategy. First, given a dataset of chillers of varying capacities, each of which operates at a fixed LWT and varying loads, iChill meta-learns a model for power prediction. Specifically, for an unseen target chiller, the meta-learned model is re-trained with known LWT to predict power at unseen LWT. Second, given the configuration of a chiller plant and a cooling load profile, iChill learns to schedule individual chillers by jointly deciding the ON/OFF status and LWT; using deep reinforcement learning (DRL). We train and evaluate iChill in a simulated environment with real-world data from a chiller plant of 22 chillers. Specifically, we compare iChill's (1) meta-learned power model with regular transfer learning; and (2) DRL scheduling with multiple baselines including CLC and an oracle model-based predictive control (MPC) strategy with perfect knowledge. We find that iChill's (1) meta-learning improves over transfer learning by up to 15.5%; and (2) DRL scheduling saves 11.5% energy over CLC and is comparable with oracle MPC (12% over CLC). Finally, off-line pre-training of iChill's DRL on the meta-learned chiller models reduces the need for real-world training experimentation by 11x from 3 years to 96 days.