OpenCEM: Multimodal photovoltaic and battery microgrid dataset with context records and sky imagery

We present a multimodal dataset from an on-campus photovoltaic and battery microgrid in Shenzhen, China. The release contains 15,543,099 rows from two hybrid inverters, including electrical measurements, operating states, settings, cumulative statistics, specifications, and fault-history arrays. A context table contains 25,178 interval records derived from manual entries, workstation workloads, system events, shell-history summaries, public calendars, and reproducible operating annotations. Each context record distinguishes its applicability interval from the time at which it was entered or derived. A roof-mounted camera contributes 82,657 upward-facing 1920×1920-pixel JPEG sky images at a nominal one-minute cadence. The resulting record links device state, local sky conditions, and availability-aware operational context without reducing them to a preselected modelling target. The synchronized modalities support studies of photovoltaic generation, battery operation, load dynamics, system identification, forecasting, multimodal representation learning, and context-aware control. The release is distributed as a compressed SQLite database and a separate image archive with a column dictionary, manifests, checksums, validation reports, and reusable processing code.