PJM Peak Load Forecasting: A Probabilistic Framework for 5CP Risk Management

Executive Summary

PJM summer 5CP (Coincidental Peak) forecasting is a peak-ranking problem in which the objective is to determine whether a future qualifying daily peak could enter the five highest system peaks of the summer. Because the final CP set is determined from observed RTO load, forecasting requires more than predicting high-demand hours; it requires dynamic comparison against the evolving CP position and explicit treatment of forecast uncertainty.

This case study describes QR Forecaster’s approach to this problem. The system generates hourly PJM RTO load forecasts, extracts the maximum forecasted load for each future day, and evaluates each daily peak against the applicable CP screening level. Before five CPs are established, QR applies a proprietary Base CP Threshold. Once five CPs exist, the fifth-highest observed peak becomes the dynamic CP Boundary for subsequent forecasts.

QR then evaluates the probabilistic daily peak distribution to estimate the probability that a forecasted peak will exceed the applicable CP level. The resulting CP-risk signal can be monitored through a seven-day reporting interface, distributed through automated email and messaging alerts, or integrated into client energy-management and demand-response systems.

The methodology is evaluated using load and peak forecast error, CP-event precision and recall, and probabilistic measures such as Brier Score and calibration. For operational CP forecasting, recall is particularly important because missed CP events can eliminate advance opportunities for peak-management actions. The resulting framework provides clients with a continuously updated measure of forecasted peak magnitude, CP position, probability, and lead time to support operational decision-making.

1. PJM Load and the Summer Peak Problem

PJM coordinates the operation of the regional transmission system and administers wholesale electricity markets across a large multi-state footprint. System demand changes continuously with weather, time of day, calendar effects, customer behavior, and other factors. During periods of extreme summer heat, air-conditioning demand can increase simultaneously across large parts of the PJM footprint, producing exceptionally high RTO load. We also use RTO for PJM RTO.

This makes summer peak forecasting different from ordinary hourly load forecasting.

A standard load forecast asks:

What will PJM load be at a particular future hour?

A 5CP forecasting application asks:

Could the peak on a future qualifying day become one of the five highest peaks of the summer?

The second question requires both load forecasting and dynamic peak scoring.

2. What Are PJM Summer 5CPs?

We first define the exact rules PJM uses to identify which 5 hours in a year count as the 5CP.

2.1. Five highest daily peak loads

Unrestricted (or unadjusted) means the raw, total physical electricity drawn from PJM’s grid. before PJM applies post-event mathematical adjustments, backcasts, or curtailment offsets.

PJM looks at the total electrical demand (in MW) across its entire region every hour of the summer. Out of the thousands of hours in a summer, PJM selects the 5 specific hours where total grid Unrestricted demand reached its absolute highest points.

The peak is measured at the entire regional system level, not at the local utility or state level. PJM adds up demand across all 13+ states in its territory (from Illinois to New Jersey). The single hour where this massive combined total hits its peak is a system peak hour.

PJM only looks for five coincidental peak (5CP) hours during these 4 summer months. Even if an extremely cold winter day in January records higher total demand than a mild June day, winter hours are excluded from the 5CP calculation.

The calculation is subject to a maximum limit of one peak hour per calendar day. This is a critical rule to prevent double-counting single heatwaves:

• If Tuesday, July 15th has extreme heat and records huge loads at 3:00 PM, 4:00 PM, and 5:00 PM, only the single highest hour of that day (e.g., 4:00 PM) can qualify.
• The other 4 peak hours must come from 4 completely different calendar days during the summer.

In summary, to find the 5CP hours, PJM ranks the single hottest/highest-demand hour of every summer day from June 1 to September 30, picks the top 5 days with the highest peak hours, and uses those 5 specific clock-hours to bill utilities and large industrial customers for their transmission and capacity charges.

2.2. Coincident Peak (CP) & Settlement Mechanics

For an entity E, a facility, Load Serving Entity (LSE), or zonal sub-region, the Coincident Peak (CPE ) active power is the power demand (in MW) recorded at a specific meter, during the exact interval in which PJM regional transmission system reaches one of its five aggregate maximum demand.

Where:

• CPE = Coincident Peak active power demand of entity E (MW)
• LE(t) = Active power demand of entity E at time interval t (MW)
• t system_peak = Time interval of one of 5 maximum aggregate PJM system load CP

The 5-Coincident Peak of the asset E, 5CPE, is the arithmetic mean of the entity E’s active power demand (in MW) measured during the 5 discrete CP intervals recorded across PJM RTO footprint during the summer period (June 1 through September 30), subject to a maximum limit of one peak hour per calendar day.

= (1 / 5) * ( LE (t system_peak_1 ) + … + LE (t system_peak_5 ) )

In reality, the 5CPE is calculated bottom up. Entity E has many consumers j

The 5-Coincident Peak for an individual customer j is

1 / 5) * ( Lj (t system_peak_1 ) + … + Lj (t system_peak_5 ) ) * Zonal Loss Factorj * Weather Normalization Scaling Factork

where:

• Li (t system_peak_k): Customer j’s metered demand during 5CP peak hour k = 1, …, 5.
• Zonal Loss Factorj: Voltage-level loss adjustment (Primary, Secondary, Transmission) for customer j.
• Weather Normalization Scaling Factork: Hourly EDC zonal factor scaling actual weather to design peak weather.

Finally, the total 5 CP for the LSE E is 5CPE (MW) = Sum over of all consumer j (Customer 5CPj ).

Indeed,

LE (t system_peak_k ) = Sum over of all consumer j Li (t system_peak_k )     (2)

Substitute (2) in (1)

Sum over of all consumer j (1 / 5) * ( Li (t system_peak_k ) + … + Li (t system_peak_k ) ) =

Sum over of all consumer j 5PLCj

2.3 PJM Tariff & Cost Allocation Impact

2.3.1 Cost Allocation from PJM to LSE

In PJM, a Load Serving Entity (LSE) E incurs two major fixed wholesale demand charges based on its customers’ performance during the five summer peak hours (5CP). PJM passes these payments directly to local Transmission Owners (such as AEP, ComEd, or Dominion) to cover their FERC-approved annual revenue requirements. PJM acts strictly as a net-zero clearinghouse, collecting capacity and transmission fees from LSEs and disbursing them directly to resource generators and transmission utilities.

1. Capacity Costs (PLC)

To ensure long-term grid reliability, PJM assigns every LSE, E, or customer a Peak Load Contribution (PLC) based on their 5CPE load, adjusted for local weather normalization and line losses. The LSE E pays PJM a daily capacity fee calculated as:

PJM passes 100% of these collected funds directly to Grid Owners that cleared the Reliability Pricing Model (RPM) auction.

1. Transmission Costs (NITS)

To fund regional grid infrastructure, PJM calculates the LSE’s percentage share of total zonal demand during 5CP hours. The (LSE) E pays a monthly Network Integration Transmission Service (NITS) fee calculated as:

2.3.2 Cost Allocation from LSE to Consumers

A Load Serving Entity (LSE) recovers its wholesale PJM capacity and transmission liabilities by passing these charges down to retail customers (index j), particularly commercial and industrial consumers who actively manage and curtail load during predicted peak hours.

1. Daily Capacity Charge Allocation (RPM Market)

The LSE bills an individual consumer for wholesale capacity on a daily basis using the customer’s assigned Peak Load Contribution tag (PLCj ) and the PJM daily auction clearing price.

Where

• Customer PLCj is a fixed megawatt (MW) tag assigned to customer account j by the local utility (EDC) for the planning year (June 1 – May 31). It is derived from the customer’s raw 5CP performance, weather-normalized, and adjusted for distribution losses. requirement for the entire planning year (June 1 through May 31).
• Zonal UCAP Price: The PJM Reliability Pricing Model (RPM) capacity market clearing price for the zone (in $/MW-day).

1. Transmission Cost Allocation (NITS)

The LSE bills an individual consumer for wholesale transmission on a monthly basis based on their proportionate share of the LSE portfolio’s peak load during the five system peak hours:

= Consumer Peak Share * Total LSE Monthly NITS

where,

• Consumer Peak Share (%) = Consumer 5CPj (MW) / Total Portfolio 5CPE (MW)
• Consumer 5CP is the Customer 5CPj average demand during the 5 summer peak hours (in MW).
• Total Portfolio 5CPE (MW) is the 5CP of the entity E.
• Total LSE Monthly NITS Bill: The total transmission dollar bill the LSE receives from PJM (in $/month).

3. The Challenge of Forecasting a CP

The primary operational objective in Coincident Peak (CP) forecasting is balancing false positive alerts against true positive captures.

• Over-predicting potential CP hours creates unnecessary operational disruptions, unwarranted asset curtailments, and customer fatigue for Load Serving Entities (LSEs) and large end-use consumers.
• Conversely, failing to capture actual CP hours (false negatives) directly exposes entities to unmitigated Peak Load Contribution (PLC) and Network Integration Transmission Service (NITS) tariff liabilities for the subsequent planning year.

This forecasting challenge is compounded by the seasonal structure of the PJM market rules:

1. Seasonal Reset & Rolling Horizon: On June 1 of each planning year, the baseline threshold of the five CP hours resets to zero. Early-season system peak loads recorded in June must be evaluated without knowing the absolute peak demand magnitudes that will occur in late July or August.
2. Threshold Ambiguity in Daily Load Forecasts: Absolute high-load forecasts do not automatically guarantee a CP event. A relatively high summer load forecast early in the season may ultimately fall outside the final seasonal top five, while a modest forecast during an unexpected late-summer heatwave may qualify.
3. Alert Fatigue Mitigation via Scoring Systems: Simply triggering demand-response or curtailment protocols for every elevated daily load forecast generates prohibitive operational overhead and financial disruption. To optimize action thresholds, a probabilistic scoring mechanism must be implemented to rank daily peak projections and assign an objective CP likelihood score before issuing actionable alerts.

4. Calculating CP Under Uncertainty

The industry standard methodology for predicting Coincident Peak (CP) hours uses a probabilistic framework to assign a likelihood to a CP event, analogous to a weather forecast predicting a 60% probability of precipitation.

The QuantRisk QR ALPS (AI Load Prediction System) approaches CP identification as a forecasted peak-scoring problem under uncertainty. The model produces an hourly probabilistic AI forecast for PJM RTO load over a 7-day-ahead (7-DA) horizon, extracts the daily peak hour for each qualifying day, and evaluates the probability of that peak displacing an established CP threshold.

4.1 Probabilistic AI Load Forecasting

QR ALPS generates a 100-percentile probabilistic RTO load forecast for each hour across the 7-DA forward horizon, denoted as RTO Load Pxx (d,t), where Pxx represents the percentile rank (e.g., P50 for the median forecast).

For forecast day d, using the median forecast RTO Load P50 (d,t), the system identifies the highest forecasted RTO load for the day

Daily Peak Forecast (d) = maxt ( RTO Load P50 (d,t) )

The resulting Daily Peak Forecast (d) series is then evaluated against the current state of the 5CP baseline set.

4.2 CP Screening Before Five CPs Are Established

Early in the summer planning period (beginning June 1), prior to five qualifying seasonal peaks being established, a screening mechanism is required to discard daily peak forecasts that have no realistic chance of becoming a 5CP event, avoiding unwarranted curtailment alerts.

QuantRisk calculates an initial Base CP Threshold for the planning year derived from historical zonal peak distributions. Only days where Daily Peak Forecast (d) exceeds the Base CP Threshold are retained as CP candidates for probabilistic evaluation.

4.3 CP Screening After Five CPs Are Established

Once five qualifying peak hours have occurred during the summer period, the screening reference transitions from the static Base CP Threshold to the dynamic CP Boundary, defined as the lowest observed peak hour in the current top five set (CP5):

For any future forecast day d, if Daily Peak Forecast (d) exceeds CP5, the day is flagged as a potential 5CP candidate. Forecast peaks falling below CP5 are mathematically incapable of entering the top-five set.

When an actual qualifying peak occurs that exceeds CP5, it enters the set, displacing the previous lowest peak, and CP5 is updated upward to establish the new CP Boundary.

As an example, a forecast below the boundary is not capable of entering the top-five set.

When a new qualifying peak displaces CP₅, the ranking is updated and the resulting lowest value becomes the new CP Boundary.

4.4 CP Probabilistic Scoring

A candidate daily peak is not evaluated solely from its point forecast. QR ALPS generates a probabilistic distribution for the forecasted load and evaluates the portion of forecast percentiles that exceeds the applicable CP screening level.

The Daily Peak Forecast (d) occurs at a time t*, represented by (d,t*)

Daily Peak Forecast (d) = max ( RTO Load P50 (d,t) ) = RTO Load P50 (d,t*)

The peak load (d,t*) is a qualified CP if

RTO Load P50 (d,t*) > Current CP Boundary CP5.

Assigning a likelihood to this event uses QR ALPS AI probabilistic forecasts:

QR ALPS generates 100 percentile forecasts for RTO Load (d,t*), namely

RTO Load P1 (d,t*), … , RTO Load P99 (d,t*).

To assign a likelihood score to this event, QR ALPS evaluates its 100 percentile forecasts at hour (d,t*), spanning RTO Load P1 (d,t*) through RTO Load P100 (d,t*). The CP Probability Score is defined as the percentage of percentile bands that strictly exceed CP5:

This probability reflects the true risk that the future daily peak will surpass the current 5CP threshold, accounting for model variance and weather uncertainty.

4.5. Seven-Day CP Score Prediction

It is best industry practice to publish 5CP score predictions for a 7 DA horizon. As QR ALPS provides consultant 7-DA probabilistic load forecasts, the process described above is applied verbatim across the 7 DA forward horizon. For each forecast day, the system produces:

The resulting seven-day output provides, for each forecast day, the expected peak load, the applicable CP comparison level, and the estimated probability of exceeding that level. As observed summer peak information changes, the CP Boundary is updated and the forecast probabilities are recalculated accordingly.

This allows CP risk to be treated as a continuously updated probabilistic ranking problem rather than a binary prediction of whether a future load value will be high.

One quick word about the industry standard 7 day-ahead horizon. The most important feature in AI load forecasting is weather, and going forward beyond a few days, weather forecasts lose their precision and are not actually available from weather data providers.

Standing on July 24, The Forecast will look like this.

5. Operational CP Risk Monitoring and Alerts

The last aspect is the publication of 7-DA CP predictions. QR ALPS converts the CP forecasting output into a daily operational monitoring system through a client-facing reporting dashboard and automated alerts. The objective is to give users a clear view of upcoming CP risk, sufficient lead time to evaluate operational actions, and a consistent way to monitor how that risk changes as new forecasts become available.

5.1. Daily CP Risk Dashboard

The QR reporting interface provides a rolling seven-day view of CP risk. For each forecast day, clients can monitor the expected RTO peak, applicable CP Boundary, estimated CP probability, candidate status, and forecast lead time. This allows users to identify which upcoming days require attention without manually processing hourly load forecasts.

The dashboard can also display the current five CPs and the resulting CP Boundary, providing direct visibility into the position of the current summer peak set. As new forecasts and observed load data become available, the dashboard is updated so that users can monitor changes in both forecasted peak risk and the underlying CP Boundary.

5.2. Automated CP Alerts

For near-term operational decisions, QR can provide automated email and messaging alerts when a forecasted event reaches a defined CP-risk condition. For example, a next-day notification can identify that the forecasted peak has a specified probability of exceeding the applicable CP level.

A typical alert can include:

• Forecast date and hour of the CP
• Current CP Boundary
• Estimated CP probability
• Forecast horizon
• Risk classification

This allows operational teams to receive the relevant information directly rather than continuously monitoring the reporting interface.

5.3. Forecast Updates and Risk Evolution

CP risk is not static. As the forecast horizon shortens and new weather and load information become available, the estimated daily peak and its probability can change. QR can therefore refresh the CP-risk output as new forecast information becomes available, allowing clients to distinguish between an event that is consistently high-risk and one whose probability is changing materially as the event approaches. The reporting system can retain the latest forecast state while updating the associated CP probability and candidate status.

5.4. Configurable Alert Thresholds

Different clients may have different operational constraints and tolerance for false-positive alerts. QR can therefore support configurable alert thresholds, allowing notifications to be triggered only when CP probability reaches a client-defined level. This creates a more targeted alerting workflow. A client requiring early warning can use a lower probability threshold, while a client seeking to minimize unnecessary operational interventions can use a higher threshold. The underlying forecast remains the same; the alerting layer determines when that information becomes actionable for the individual user.

5.5. Integration with Client Systems

For clients that already operate energy-management, demand-response, scheduling, or internal analytics systems, CP-risk information can also be made available through data feeds or API-based integration. Rather than requiring users to manually transfer information from the QR dashboard, forecast peaks, CP probabilities, CP boundaries, and related risk indicators can be incorporated into existing operational workflows. This enables CP risk to become an input to automated or semi-automated processes such as demand-response preparation, flexible-load scheduling, storage planning, or internal peak-management applications.

The resulting system provides clients with three complementary ways to use the forecast: monitor CP risk through the daily dashboard, receive automated alerts when predefined conditions are met, and integrate the forecast directly into existing operational systems.

6. Operational Applications of CP Risk Forecasting

The value of CP forecasting is realized when the forecast is incorporated into day-to-day operational decisions. The appropriate use of the forecast depends on the client’s load profile, flexibility, commercial arrangements, and response capabilities, but the same CP-risk indicators can support several operational workflows.

6.1. Utilities and Load Serving Entities

Utilities and LSEs can use the seven-day CP-risk outlook to identify periods requiring closer operational attention. Rather than responding only after a system peak has occurred, planning teams can monitor forecasted peak levels and CP probabilities in advance and coordinate potential load-management activities.

Potential applications include:

• Preparing demand-response programs for high-risk days
• Coordinating customer peak-management communications
• Reviewing expected capacity exposure
• Prioritizing operational resources for potential peak events
• Monitoring changes in CP risk as the forecast horizon approaches
• Supporting internal peak-management and planning decisions

The forecast can therefore serve as an additional operational input alongside weather, load, and customer-level information.

6.2. Large Commercial and Industrial Consumers

For large energy consumers, a potential CP event can provide an advance signal to evaluate flexible-load opportunities. Where operationally and economically feasible, customers can use the forecast to review actions such as HVAC adjustments, production scheduling, battery dispatch, thermal storage, or participation in demand-response programs.

The key advantage is lead time. A CP-risk notification received before the potential event gives an energy manager an opportunity to evaluate the cost of reducing or shifting load against the potential value of doing so.

6.3. Demand-Response Providers and Aggregators

Demand-response providers and aggregators can use CP probability and forecast lead time to prioritize resources ahead of potential peak events. Rather than treating every elevated load forecast as an activation opportunity, the CP-risk signal can be used to rank upcoming days and focus preparation on events with a higher probability of becoming significant CP candidates.

This can support resource availability checks, customer notifications, response planning, and coordination of flexible assets before the expected peak period.

6.4. Energy Managers and Consultants

Energy managers and consultants can use the QR outputs as a common monitoring layer across multiple facilities or customer accounts. The combination of forecasted peak, CP Boundary, CP probability, and forecast horizon provides a consistent basis for reviewing upcoming peak risk and determining which customers may require attention. The same information can also be incorporated into recurring daily reports, internal operating procedures, or customer-specific peak-management workflows.

7. Forecast Evaluation and CP Detection Performance

Evaluation of QR’s CP forecasting system should separate load-forecast accuracy, CP event detection, and probability quality. A model can produce an accurate point forecast while still performing poorly at identifying the relatively small number of daily peaks that ultimately become CP events. CP-specific evaluation is therefore required in addition to conventional load-forecast metrics.

7.1. Load Forecast Accuracy

For the underlying RTO load forecast, standard error metrics can be calculated over the forecast horizon. Mean Absolute Error (MAE) is defined as:

where N is the number of evaluated settlement intervals.

For CP applications, interval-level MAE should be complemented by daily peak error, since the maximum load of each day is the quantity used by the CP screening process. A corresponding daily peak error can be expressed as:

and:

where D is the number of evaluated days.

7.2. CP Candidate Detection

CP detection should be evaluated as a binary classification problem. For each forecast day, the model’s CP candidate signal can be compared with the eventual observed outcome:

• True Positive (TP): A day identified as a CP candidate that subsequently qualifies as a CP event.
• False Positive (FP): A day identified as a CP candidate that does not subsequently qualify as a CP event.
• True Negative (TN): A day not identified as a CP candidate that does not subsequently qualify as a CP event.
• False Negative (FN): A day not identified as a CP candidate that subsequently becomes a CP event.

Two fundamental detection metrics are then:

Precision measures how frequently an issued CP signal corresponds to an actual CP event. Recall measures how effectively the forecasting system captures the CP events that actually occur.

7.3. Why Recall Is the Primary Metric

For CP forecasting, recall is generally the more important detection metric because a false negative represents an actual CP event that was not identified sufficiently in advance. This can remove the opportunity for a client to evaluate load reduction, demand-response activation, storage dispatch, production rescheduling, or other peak-management actions.

A false positive has a different consequence: it may result in an unnecessary operational review or intervention. Therefore, precision remains important because a system that generates excessive false positives can create alert fatigue and unnecessary operational costs. However, increasing precision by making the model excessively conservative can reduce recall and cause the system to miss genuine CP events.

The operational objective is therefore not to maximize precision independently. It is to maintain 100% recall while controlling false-positive frequency. In practical terms, the alerting threshold should be selected based on the relative cost of:

where the exact cost relationship is client- and application-dependent.

This also means that CP detection should be evaluated at the same probability threshold used for operational alerts. Changing the probability threshold changes the balance between TP, FP, and FN and therefore changes both recall and precision.

7.4. Probabilistic Forecast Evaluation

Because QR produces a CP probability rather than only a binary candidate signal, the probability estimates themselves should also be evaluated.

For each forecast day d:

where:

• p(d) = predicted probability of the CP event
• o(d) = observed outcome, equal to 1 when the CP event occurs and 0 otherwise
• D = number of evaluated forecast days

A lower Brier Score indicates better probabilistic accuracy.

8. Conclusion

PJM 5CP forecasting is fundamentally a peak-ranking problem under uncertainty, rather than a conventional hourly load-forecasting problem. The objective is not simply to predict when PJM RTO load will be high, but to determine whether the peak of a future qualifying day has a realistic probability of entering the five highest system peaks of the summer.

QR ALPS addresses this by combining 7-day-ahead probabilistic RTO load forecasting with dynamic CP scoring. The system first generates hourly load forecasts and identifies the daily peak for each forecast day. Before five CPs are established, a proprietary Base CP Threshold is used to screen potential candidates. Once five qualifying peaks exist, the fifth-highest observed peak becomes the CP Boundary, which is continuously updated as the summer peak set changes.

The forecasted daily peak is then evaluated using the probabilistic load distribution rather than relying only on the median forecast. This produces a CP probability score representing the likelihood that the forecasted peak will exceed the applicable CP level. Because the boundary and forecast distribution both evolve over time, CP risk can be reassessed continuously as new load and weather information becomes available.

For operational users, this converts a difficult seasonal ranking problem into a forward-looking risk signal. Utilities, LSEs, large commercial and industrial consumers, demand-response providers, aggregators, and energy managers can use the resulting forecasted peak, CP Boundary, probability score, and lead time to determine whether operational action is warranted.

The forecasting system can be evaluated at three complementary levels: underlying load accuracy, CP-event detection performance, and probabilistic forecast quality. Recall is particularly important because missing an actual CP event can eliminate the opportunity for advance peak-management action, while precision remains necessary to limit unnecessary interventions and alert fatigue.

QR ALPS therefore does not replace PJM’s eventual determination of the five coincident peaks. Instead, it provides a forward-looking probabilistic forecasting layer that estimates CP risk before the final summer peak set is known. The objective is to give market participants sufficient lead time and quantitative risk information to evaluate appropriate operational responses as the summer peak develops.

9. Explore QR AI Load Forecasting Solutions

Discover QR’s AI-powered load forecasting solutions, explore technical architectures and additional case studies, or access our live interactive forecasting dashboards to see how probabilistic load forecasting can support electricity market and grid planning applications.

Energy Load Forecasting Solutions.

For inquiries or comments, please contact [email protected]. Our team is available to discuss your forecasting requirements and potential applications of QR Forecaster.