Definition
Measures the degree to which AI dispatch recommendations remain cost-optimal under the current grid regime compared to actual least-cost dispatch determined by post-hoc settlement analysis.
The Measurement Gap
Traditional dispatch operations track heat rate, capacity factor, and fuel cost per MWh at the unit level. These metrics tell you whether individual generation units are running efficiently. They do not tell you whether the unit commitment and dispatch sequence recommended by the AI was the correct economic choice given the actual market conditions that materialized. The gap between "the unit ran efficiently" and "the AI told us to run the right units in the right order" is precisely where DRAI operates.
The specific fall-short: EMS recommendation logs capture what the AI advised. Settlement systems capture what the actual least-cost dispatch would have been, in hindsight. No existing metric joins these two data streams to compute the AI's dispatch quality as an economic ratio. Operators know when they deviated from the AI's recommendation, but they have no systematic way to measure whether the AI itself was right.
The exact operational decision DRAI enables:
whether to retrain or recalibrate the dispatch AI, and which regime-change factors (DER penetration, storage cycling behavior, load shape shifts, renewable intermittency) are driving the largest divergence from optimal.
What goes wrong without this information:
the dispatch AI ages silently. Its recommendations grow progressively more expensive relative to optimal as the grid regime evolves. Operators may sense something is off but lack a quantified signal to trigger model review. The cost of sub-optimal dispatch accumulates in fuel and market exposure without attribution.
How AI/ML makes this measurement possible:
settlement systems now produce interval-level LMP data that enables ex-post optimal dispatch computation. Pairing this retroactive calculation with the AI's logged recommendations creates a performance measurement capability that wasn't feasible before real-time settlement data was widely available.
Formula
DRAI = Σ(Cost_AI_recommended_i) / Σ(Cost_actual_optimal_i)
Variables:
Cost_AI_recommended_i: production cost of the dispatch action the AI recommended for interval i, whether or not the operator followed it ($/interval). Source: EMS dispatch advisory recommendation log
Cost_actual_optimal_i: production cost of the ex-post least-cost dispatch for interval i, computed from settlement LMPs and actual system conditions ($/interval). Source: ISO/RTO market settlement system
Calculation Frequency: Rolling 30-day periods, computed daily. Aggregated at the interval level (5-minute or 15-minute dispatch intervals depending on market structure). Minimum 500 dispatch intervals per measurement period.
Edge Cases:
If the operator overrode the AI recommendation, use the AI's recommended action in the numerator. This metric evaluates model quality, not operator judgment. During declared reliability emergencies or forced-outage periods where dispatch was constrained by reliability rather than economics, exclude those intervals from the denominator.
Output:
Dimensionless ratio. 1.0 = AI dispatch aligns perfectly with optimal. Values above 1.0 represent the cost premium of the AI's recommendations above least-cost dispatch. A DRAI of 1.05 means the AI's recommendations cost 5% more than optimal on average. Healthy target: DRAI below 1.03. Model review warranted above 1.08. Retrain trigger above 1.10 sustained over two consecutive 30-day periods.
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Should we retrain or recalibrate the dispatch AI, and which regime-change factors (DER penetration, storage behavior, load shape shifts) are driving the largest divergence?