Definition
Measures which electrification assumptions (EV adoption rate, heat pump conversion rate, data center load growth, behind-the-meter solar penetration) contribute the most error to the feeder-level load forecast.
The Measurement Gap
Traditional load forecasting validation compares predicted aggregate load to actual aggregate load at the substation or planning area level. This comparison produces an error statistic (MAPE, RMSE) that tells planners whether the forecast was accurate but nothing about why it was inaccurate. In a multi-assumption model with EV adoption, heat pump penetration, solar growth, and commercial development all running simultaneously, the aggregate error is the sum of all assumption errors: some may be canceling each other out, and the planner has no way to know.
The specific failure point is capital investment misalignment driven by the wrong assumption. If the load forecast overestimates EV adoption by 30% in a planning zone, the utility may invest in substation capacity that sits underutilized for a decade. If it underestimates data center load growth, feeders overload. In both cases, the error may not be visible until the capital decision has been made and construction is complete. LFAAI makes the assumption error visible before the next planning cycle.
LFAAI enables the specific decision of which data source needs to be improved or replaced. If LFAAI consistently shows that EV adoption assumptions contribute 45% of forecast error, the utility should invest in utility-owned charging data, DMV registration matching, or distribution-connected EVSE counts: targeted improvements to the highest-error assumption rather than broad model retuning.
Without this measurement, planning teams retune models based on aggregate accuracy statistics while the highest-error assumptions continue to drive infrastructure decisions. Capital is misallocated because the model improvement effort is unfocused.
Retrospective AI decomposition using Shapley values makes multi-assumption attribution computationally feasible at the feeder level, which was not practical with spreadsheet-based planning models.
Formula
LFAAI(a) = |Error_attributable_to_assumption_a| / |Total_forecast_error|
Variables:
- Error_attributable_to_assumption_a: contribution of assumption category a's deviation from reality to total forecast error at feeder or planning zone level (MW). Derived from sensitivity decomposition or Shapley value analysis of the planning model; Source: load forecasting/planning model sensitivity coefficients
- Total_forecast_error: predicted peak load minus actual peak load at the feeder or planning zone level (MW). Source: AMI/SCADA metering vs. planning model output
- Assumption categories (a): EV adoption rate, heat pump conversion rate, behind-the-meter solar penetration, data center/large commercial load growth, organic residential growth. Source: planning model input documentation matched to actuals from interconnection records, DMV data, building permits, CIS
Calculation Frequency: Annual retrospective, computed 12–18 months after the planning cycle to allow actuals to accumulate
Output: Percentage per assumption category, summing to 100% of total forecast error. Any assumption contributing above 40% of total error warrants targeted data improvement for that input source. Track LFAAI across multiple planning cycles to identify whether high-attribution assumptions are persistent failures or one-time misses.
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
When the load forecast is wrong, which assumption broke, and which capital investment decisions were based on the faulty assumption?