Firm-Power Availability Forecasting

L7 forecasts how much firm power each unit can guarantee over the horizon by fusing twin health estimates, maintenance plans, and fuel availability.

Forecasting availability, not weather

A fusion source is not intermittent like wind or solar, so forecasting is not about weather; it is about the machine. The question is: over the next hours to weeks, how much export can this unit firmly commit to, and with what confidence? L7 answers by combining the L3 twin's health and confidence estimates, the maintenance service loop's planned outages, and fuel (helium-3) availability into a probabilistic availability forecast.

Inputs to the forecast

• Twin confidence: per-module health and remaining-useful-life on magnets, DEC train, plug
• Planned outages: scheduled maintenance windows from the service loop
• Fuel supply: committed helium-3 inventory and inbound deliveries
• Historical de-rate: distribution of unplanned trips from fleet learning

# probabilistic availability for unit i over horizon h
# returns firm level p_firm with confidence 1-alpha
def firm_forecast(i, h, alpha=0.05):
    base   = twin_capacity(i, h)                 # nominal export
    derate = rul_derate(i, h)                     # from RUL models
    outage = planned_outage_mask(i, h)           # service loop
    fuel   = min(1.0, he3_available(i,h)/he3_need(i,h))
    samples = monte_carlo(base, derate, outage, fuel, trip_dist(i))
    p_firm  = quantile(samples, alpha)            # conservative commit level
    return p_firm, confidence(samples)

The forecast is deliberately conservative: the firm level is a low quantile of the sampled availability distribution, so committing to it is safe even when several risks land together. The expected level is reported separately, but only the firm level is offered to the grid as a guarantee.

Forecast errors feed back. When actual availability diverges from forecast, the residual updates the trip distribution and de-rate models via fleet learning, so the forecast sharpens as the fleet accumulates operating hours. Early units contribute the most learning; by NOAK and BOAK the availability distribution is tighter.

Today the forecast runs on twin-simulated operation and historical device analogues; it reports availability as design-stage until real units exist. This keeps the firm-power promise honest: no availability is claimed for hardware that has not been built and run.