What Joule does.

The cascade decides what is theoretically optimal every 5 minutes. The safety guardrails decide what Joule is actually allowed to recommend, so that an aggressively-tuned cost model never produces an operationally-damaging instruction.

Anti flap protection. Minimum hold times of 15 minutes apply to every recommended source switch. Diesel cool down periods prevent re-cycling generators that have just stopped. The system cannot recommend a switch that would damage equipment by rapid cycling, even if the math says to.

Savings thresholds. No switching unless the projected saving exceeds a configurable threshold (default $0.50/hour). Joule ignores micro-optimisations that would generate operational churn for trivial dollar gain. The threshold is per site and tunable.

Hysteresis bands. A 5%-up / 10%-down logic prevents switching on partly cloudy days where solar fluctuation alone would otherwise trigger flapping. The platform needs a clear benefit signal in either direction before it changes anything.

Deterministic and auditable. Every recommendation includes a human readable reason and a confidence score. There is no "black box" decision. Operators can replay any decision against the inputs that produced it, for any site, for any cycle. Compliance and audit reproduction is a query, not a project.

The cascade decides what to do right now. The machine learning layer decides what to do across the next 24 hours, given what Joule already knows about the site, the local weather forecast, and the demand pattern.

Solar generation forecasting. Combines weather API data with the site's historical efficiency curve to predict the exact solar output curve for the next 1 to 24 hours. Cloud cover impact analysis. Time of day bell-curve logic. Historical shading patterns. The forecast is updated hourly as the weather forecast updates.

Load and demand forecasting. Learns from 30 days of stored history (time series storage, per site) to forecast power needs by hour of day and by environmental factor. Temperature-adjusted air conditioning load. Traffic-driven equipment load. Seasonal variance modelling. The forecast feeds the cascade so that the system pre-positions the battery for predicted demand peaks rather than reacting to them.

Battery dispatch optimisation. Optimal 24 hour battery charging and discharging schedule, enriched with contextual signals (thermal health, grid reliability probability for the next 24 hours, tariff schedule). The dispatch policy targets a configurable state of charge for the night ahead based on the expected grid availability and the cost of charging now versus later.

The three models run independently and feed the cascade. Where they disagree, the cascade's safety guardrails take precedence. Joule will not deviate from a safe operational path on the strength of a machine learning signal alone.

Joule is structured as four layers. Ingestion, intelligence, output, and platform. Each is independently versioned and the boundaries between them are deliberately thin, so operators can audit any decision back to its inputs. A single agent instance manages 50 or more sites concurrently. A geographic weather cache means a 10,000 site portfolio pulls roughly 300 weather API fetches per hour, not one per site per cycle.

Ingestion. Joule polls the site's Remote Monitoring System over a standard web interface every 5 minutes for 27 telemetry points spanning solar, grid, diesel, battery, and load. A weather forecast is fetched hourly, with results cached geographically. Neighbouring sites within approximately 11 km share a single fetch. Where the operator does not have a Remote Monitoring System in place, Joule can also connect directly to sensors via standard sensor messaging protocols (MQTT, CAN bus), or be deployed alongside a low cost monitoring device.

Intelligence core. The 5 case cascade evaluates source combinations against live inputs. The three machine learning models forecast solar, demand, and battery dispatch across the next 24 hours. Safety guardrails (anti flap, hysteresis, savings threshold) constrain what the system is allowed to recommend.

Output. Recommendations are emitted as cryptographically signed webhooks to the operator's NOC or SCADA endpoint. Each recommendation carries the recommended source mix, a human readable reason, a confidence score, the contributing inputs, and the dollar impact against the do nothing baseline. Anti flap and diesel cool down constraints apply at the output layer too.

Platform. A single agent instance manages 50+ sites with shared time series storage and connection pooling. A 14 endpoint REST API exposes fleet summary, site management, hot reload configuration, and historical query. Sites can be added or removed in seconds without restarting the agent. Hot reload means configuration changes (new tariff schedules, updated anti flap windows) are applied without downtime.