Low Emission Building Control with Zero-Shot Reinforcement Learning

Heating and cooling systems in buildings account for 31% of global energy use, much which are regulated by Rule Based Controllers (RBCs) that neither maximise efficiency nor minimise emissions interacting optimally with the grid. Control via Reinforcement Learning (RL) has been shown to significantly improve building efficiency, but existing solutions require access building-specific simulators or data cannot be expected every world. In response, we show it is possible obtain emission-reducing policies without such knowledge a priori–a paradigm call zero-shot control. We combine ideas from system identification model-based RL create PEARL (Probabilistic Emission-Abating Learning) short period active exploration all required build performant model. experiments across three varied simulations, outperforms an RBC once, popular baselines cases, reducing as whilst maintaining thermal comfort. Our source code available online via: https://enjeeneer.io/projects/pearl/.