Data and results for: Deep reinforcement learning control of a grid-tied microgrid with a heat pump and PCM thermal storage
Abstract
This dataset accompanies the paper "Deep reinforcement learning control of a grid-tied microgrid with a heat pump and PCM thermal storage: a comparison with rule-based control in an experimentally calibrated simulation" (under review). The code is at https://github.com/fred-stanford/pcm-tes-microgrid-rl. The study trains deep Q-network (DQN) controllers for a simulated grid-tied microgrid in Half Moon Bay, California, with a 3.84 kW PV array, a 7.2 kWh battery, a reversible air-to-water heat pump, and hot and cold phase-change-material thermal energy storage (PCM TES). It compares them with a rule-based controller under a time-of-use tariff, an export credit, demand-response events, and grid outages. Controllers were trained on 2018–2022, selected on 2023, and tested on 2024. Files:- datasets.zip: 5-minute train (2018–2022), validation (2023), and test (2024) datasets with weather, PV capacity factor, building electrical load, import price, and export credit.- weather_nsrdb.zip: NSRDB 5-minute weather for the site, 2018–2024, from which the datasets are built.- test_year_traces.zip: 5-minute test-year simulation traces for the rule-based controller and for the controller selected for each of the 10 training seeds.- results_summary.zip: summary results behind the paper's tables and figures (test-year metrics, Monte-Carlo results over 100 grid-event schedules, learning curves, validation rankings).- models.zip: all 60 trained DQN controllers (10 seeds × 3 discount factors × 2 network depths), each with its validation-selected checkpoint (Stable-Baselines3) and full settings.- README.md: file contents and how to use them with the code.- SHA256SUMS.txt: checksums. Weather data are from the U.S. National Solar Radiation Database (NSRDB, NREL). Import prices follow PG&E Electric Schedule E-ELEC, and export credits follow PG&E's Solar Billing Plan Energy Export Credit values. Building loads are synthetic.
