By Muhammad Umar. This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/grid-control-room-saudi-arabia/. DOI 10.5281/zenodo.23889768. The operator is described by class, never by name.
A national grid operator in Saudi Arabia runs the transmission grid from regional and national control rooms. When one 380 kilovolt line trips at summer peak, the control engineer needs a list with a clock on it: which protection operated and in what order, which busbars and connected operators now sit de-energized, what maintenance is already under way on the affected apparatus, which procedure governs the switching that follows, and who must approve each step. Every fact exists, spread across SCADA and EMS, the outage plan, meter data, the asset system, the protection settings, a procedure library in Arabic and English and the weather feed. Today the engineer is the join, working from memory, phone calls and window switching while the alarm flood is still arriving.
1. A boundary that only points one way
The platform never sends a control command and never writes into a system of record. It reads ten sources, all read only and all of operator provenance, through a replicated historian or DMZ feed into a separate AI platform zone. Direct read only links inside the operational network were set aside, because the segmentation controls and the operator's own separation rule make any direct path a compliance breach rather than a shortcut. A hardware data diode with forward error correction is the committed fallback wherever the segmentation review demands demonstrable one way flow; software only read only links cannot demonstrate it. The single write path anywhere in the design is the recorded confirmation or approval, written into the object model and the memory service.
2. Three adapter families, one backbone
Three adapter families own every contract. The telemetry adapter carries the replicated SCADA and EMS series into the time series store; the integration adapter carries outage plans, protection settings, procedures and weather; the ERP adapter carries asset, maintenance and meter data from the enterprise systems. The tier checks schema and units, stamps provenance on every fact as it lands, and checks citation resolution and freshness on every answer, so a draft that cites a superseded procedure or a stale feed is caught before a person reads it. Apache Kafka carries everything, partitioned by point or asset, delivered at least once with idempotent writes so a replay cannot double count an alarm. Its retained log is also the replay mechanism that validates the triage agent on the operators' own recorded alarm floods. Linuxptp with chrony, disciplined by a GNSS PTP grandmaster with holdover, puts alarms, events and telemetry in one sequence of events. The time series store is left as a class, chosen at design against the historian's export format and scan rate.
3. Sixteen objects seeded from the network model
The object model holds sixteen objects with thirteen typed links: substation, transmission line, busbar, transformer, breaker and protection relay; alarm and grid event; switching program, planned outage and shift log; operating procedure; weather cell, connected operator, control engineer and shift supervisor. Topology is seeded by the CIM export from the EMS network model rather than assembled from documents. When the line trips, one traversal starts at the alarm, reaches the line it references, the substations at both ends, the busbars behind them, the operators supplied through them, the relays that protect the line with their settings revisions, the procedures that govern restoring it and the shift logs where the last comparable event was handled. An alarm moves to Acknowledged only under the control engineer, and a switching program moves to Approved only under the shift supervisor.
4. No edge, two central tiers
There is no edge tier and there are no cameras: the platform consumes replicated data, and fast protection decisions stay in the relays and remote terminal units where they have always lived. GLM 5.3 is filed at 753 billion parameters; at FP8 that is 753 GB of weights, and a 1.2 planning factor makes 904 GB. One node of eight 141 GB class GPUs gives 1,128 GB, which leaves 375 GB beside the weights against the 151 GB the factor reserves, and that headroom is what serves concurrent requests during an alarm flood. Two nodes of the 80 GB class were set aside because they need a fabric between them, and the 180 to 192 GB class buys headroom this workload does not need. BGE-M3 at FP16, about 1.1 GB, runs on the platform zone's rack server class for Arabic and English retrieval, and its recall on the operator's own Arabic procedures is evaluated before go live. vLLM serves the weights, Kubernetes orchestrates, and updates move through Harbor as staged, signed images.
5. Four agents that draft, people who decide
GLM 5.3 drafts for four agents: alarm triage confirmed by the control engineer, the switching program approved by the shift supervisor, the handover approved by the incoming supervisor, and the event analysis prepared for the protection engineer. Nothing is trained into weights; memory loops over confirmed decisions improve each shift, because a moving model would invalidate the frozen baselines. The GLM-5.3 bespoke license allows commercial use, fine tuning and redistribution, with an attribution notice and a managed access trigger above 10 billion US dollars of group revenue that internal control room use does not reach. Holding the sizing at one node means the export license determination for the GPU node, naming this end user, is asked once, early, as a phase one checkpoint before the order is placed.
Appendix A prices the node at about 570,000 US dollars to own over three years (542,000 to 594,000), with power at the national commercial electricity tariff. Renting the same eight cards on the deepest three year commitment in AWS Middle East (UAE) is about 751,000, so ownership costs about three quarters of it, and about 2.00 million on demand, against which ownership costs a little over a quarter. Renting is a price reference only: AWS listed no Saudi region on the day the prices were read, and the grid data would cross the border.
The rollout is three phases and 18 scope items (15, 2 and 1), each closed by a gate rather than a date: a pre summer build and replay gated on the switching program agent running end to end on replayed records, one regional control center live through one summer, then extension gates. Baselines are frozen before go live, and the headline measure is trip to first correct action.
Made with CodeNinja Praxis, the CodeNinja Atoms platform for designing physical AI systems. The object model imports into Hyper Ontology. Paper, ontology and register: https://codeatoms.ai/grid-control-room-saudi-arabia/
One of the CodeNinja Atoms sovereign AI reference architectures: every published design with its sector, country, compute and three year cost.
Top comments (0)