Forecasting Load Schedules


A Background Paper for the DAWG Implementation Committee


May 23, 1996

Michael R. Jaske, Chief Forecaster

California Energy Commission

Under the design of the ISO and Power Exchange institutions to create a competitive power market and to sustain system reliability, all consumer's loads must be forecast, but on a different basis than utilities have done for purposes of system dispatch. This paper seeks to provide background on what has been done in the past, and what will have to be done differently under the proposed organization of the industry.

1. Traditional Short Run Load Forecasts

Utilities have made short run load forecasts for purposes of making unit commitments, scheduling contract resources, and dispatching resources to maintain system frequency. These forecasts have been for the total aggregation of all customers' loads, and only reflect specific customer information in special circumstances.

The techniques used for these load forecasts rely upon stability in the base of customers that the system is serving. Aggregate load data is processed using econometric and time series data processing techniques. Econometric techniques use measures of weather and other time varying phenomena to develop relationships between previous loads and a few aggregate explanatory factors. Time series modeling techniques rely upon observations of what has just been observed more than explanations what caused what has been observed. The parameters estimated in these analyses, along with short run forecasts of the explanatory variables, serve to provide short run load forecasts that serve as the basis for unit commitment and scheduling. Observations of actual system load on a quite short run basis are used to update these models on a very frequent basis, (e.g. daily or weekly).

2. Requirements of Generation Service Retailers

All of the parties serving loads, whether through bilateral contracts for larger consumers, through aggregations of many smaller consumers, or the UDC for consumers not wishing to participate in direct access, must forecast loads and provide these forecasts to one or more entities upstream of the retailer. For bilateral contract retailers, this will be directly to the ISO through a scheduling coordinator to the ISO. For aggregators, forecasts of loads will again be provided to the ISO or through a scheduling coordinator to the ISO, but the ISO needs only an aggregation of consumer's loads, not the actual loads for each consumer individually. Finally, for loads served by the UDC, the UDC provides its load forecast to the Power Exchange.

2.A How are these Load Forecasts Different in Scope?

The scope of the load forecasts provided to the ISO or to the PX are much smaller in scope than those traditionally prepared by utilities for unit commitment and dispatch purposes. Recall that those forecasts were for the entire utility customer base - the system load. The principal exceptions were major self-generation customers whose generating unit might be scheduled off for maintenance, and this event might be treated in a specific, recognizable way for scheduling utility units or contract resources.
There may be dozens or hundreds of separate schedules provided to the ISO and the PX, which collectively sum to the three IOU systems that were formerly treated as three disjoint sets of loads for the three IOU systems. Therefore, each of the load forecast schedules are much smaller in scope than the previous utility system load forecasts. Further, some of the load forecast schedules will be composed of a single or a few end-use consumers. Therefore, the methods for preparing load forecasts for these much smaller groupings of loads cannot ignore the consumer-specific features of loads, e.g. things that cause a single or a few customers to do something at specific times that cannot be explained by techniques that are based on averages of large numbers of customers.
Some aggregations of loads, and certainly the UDC load forecasts submitted to the PX, will be large enough and diverse enough that similar methods to those used traditionally by utilities would be satisfactory. However, even these load forecasts have to accommodate the fact that customers are moving into or out of various aggregations (whether that of a specific DA retailers or the UDC) all the time, and thus accounting for customer turnover in some explicit way will be necessary.

2.B How are these Load Forecasts Different in Purpose?

It is extremely important to understand that the purpose of these load forecasts will be different than those developed previously by utilities. To reiterate, the traditional load forecasts were used to make unit commitment and contract resource decisions from a relatively fixed pool of generating resources owned, operated, and controlled to varying degrees by the utility. Capital costs of these facilities and the decision to add resources (the capacity expansion question) was determined in a completely different forum than the utilization of resources.
Under the proposed market structure, however, load forecasts serve to provide a commitment to purchase power according to the specified schedule. Failure to support the load forecast does not reduce the payment that must be made to the power supply industry. In effect, you pay what you asked for if you asked for too much, and you pay for what you used if your actual is larger than what you asked for. The financial consequences of errors are much larger under the new market structure than in the previous industry structure. Previously, errors in the unit commitment process implied too many resources were brought up to spinning status, or too few were brought up, but other resources could be run somewhat harder, using more fuel than would have been optimal with perfect foresight of loads. The net result was additional fuel costs but no additional capital costs.
In contrast, under the new market structure, generators will be submitting generation bids that presumably cover both capital and operating costs. The bids may even be higher than costs justify if the generator believes that the facility can still be scheduled into the Power Exchange market. Therefore, under the new market structure, every kilowatthour of error in scheduled load forecasts versus actual loads imposes greater costs on the party responsible for the load bid.

3. Likely Difficulties with Load Forecasting for Retailers

Bilateral contract suppliers can be expected to have problems with developing load forecasts for their associated load customers simply because making such forecasts has not been developed into a refined art. Developing such forecasts obviously relies upon having and analyzing historic load patterns of the customer(s).[1] Even given the data, however, the variations from an expected pattern may be considerable, and not known to any party since historic usage was not metered or recorded in this manner in the past. Retailers will have varying levels of difficulty in making these load forecasts, with the production process of the customer being the dominant consideration.[2]
Emergent aggregators with hundreds to thousands of small customers may have the greatest difficulties since there may be no customer- specific hourly load metering historic data, and little understanding of how specific customers do or do not match up with known load research data for the broad class within which the customer had formerly been classified by the utility. Once some degree of actual metering data becomes available for the "classes" or "subgroups" of the aggregation retailer, then improvements in load forecasts can be expected.

4. UDC Load Forecasts Provided to the Power Exchange

The UDC must provide a load forecast for its full service customers to the Power Exchange, so that the Power Exchange can include these loads in the determination of generator loading order to minimize expected market price. The UDC prepares this short term load forecast very frequently on the schedule needed by the Power Exchange.
Unfortunately, the existence of DA customers mixed in with full service customers of the UDC makes the job of predicting UDC loads more difficult. The following discussion develops what would be needed for two different approaches:
a. Approach 1 presumes that the UDC makes a total regional load forecast and nets out DA loads to determine a UDC load forecast for its full service customers.
b. Approach 2 presumes that the UDC and other parties make parallel load forecasts, with direct access load forecasts going to the ISO through the scheduling coordinators, and the UDC load forecast goes to the Power Exchange.
4.A Approach 1--Netting Out DA Loads from Total Load Forecasts
Utilities are currently making short term load forecasts for use in scheduling utility powerplant operations, contract purchases, generator maintenance, etc. These forecasts represent all uses of electricity within the service area, with the exception of self- generators providing their own loads. All consumption within the region, less self-generation, is known to the utility and available for use in making consumption forecasts. The omission of self- generator loads, however, is not crucial to the load forecasting function, because self generator loads do not flow through the distribution system nor present requirements to utility generators. With the advent of DA, however, the UDC will not routinely be able to make total consumption forecasts unless the UDC continues to have access to generation usage data for all loads, not just those of its full service customers. Under the current policy decision, the UDC continues to responsible for the billing system, but D.96- 03-022 sets in motion an examination of distribution component services and it is conceivable that the UDC may lose access to consumption data for all customers at some future date.
Using historic load data for all electricity consumers in the UDC franchise service area, the UDC can prepare load forecasts for the area much like it currently does. Once the total load forecast for the franchise service area and the DA load forecast disaggregated to service area have been prepared, the computation of net UDC loads can be performed. This information is provided to the Power Exchange and used by the Power Exchange to scheduled generators in least cost merit order to determine a minimum market price to serve

UDC loads, or any others contractually served by the Power Exchange.[3]

The statistical properties of the sum of DA load forecasts will show inherently more variation than total regional loads because customers can shift back and forth between DA and UDC full service status with no penalty. A netting approach translates this variation into the net UDC load forecast. Thus all of the usual physical and economic reasons for uncertainty in load are compounded by the ability of the customer to shift status from full service to DA and back in a short period of time.[4] Since DA retailers will be required to provide schedules to the ISO, this imposes a forecasting burden on them. Errors in load forecasts will induce DA retailers to make an effort to provide reasonably reliable estimates of short term loads, because the errors result in costs that either customers have to absorb or the aggregator's profits must cover. In Approach 1, the UDC needs DA load forecasts categorized by UDC service area, the DA suppliers would have an obligation to provide their best load forecast information to the entity that makes the net UDC load forecast, otherwise the UDC's customers would be paying for mistakes made by the DA retailer.

The following questions exist for Approach 1:

a. Should the UDC net out the DA loads in its franchise service area? If the UDC has this obligation, how will DA loads be provided to the UDC?

b. Should the PX net out the DA loads applicable to a UDC service area to obtain a UDC full service customer load forecast? If the PX has this obligation, how will DA loads be provided to the PX?

c. Are there competitive reasons for the UDC or PX to not be provided loads of aggregators and bilateral direct access customers which eliminate netting as a feasible approach?

d. What responsibility for costs would UDC full service customers bear when errors in its load forecast were caused by errors in DA load forecasts?

4.B Approach 2--Parallel Preparation of Load Forecasts

Approach 2 presumes that UDCs are notified that specific customers have entered into DA arrangements (which would be required so as to be able to offer them an appropriate tariff for remaining distribution services, so this is not a new requirement just for load forecasting purposes), and the UDC is able to make use of this information in making load forecasts for the UDC's full service customers. Thus, the UDC does not have a forecast prepared by netting out forecasts made by others, but rather focuses on preparing a forecast for the customers for which it continues to be the energy supplier.

In order to do this, the UDC would have to adapt its current load forecasting techniques to recognize, perhaps, some broad classes of customers who might be expected to behave homogeneously, and it continuously adjusts these class load forecasts for the number of customers for which it has an energy supply responsibility. One could imagine that some time series techniques would continue to be used, but at a customer class level, and that through a multiplier or as an explanatory variable the aggregate load of this customer class was adjusted as numbers of customers in that class increase or decrease.

Accomplishing this result might require greater reliance upon customer class load research data in real time, rather than in the very much delayed way in which load research data is now collected, processed, and made available for use for analytic purposes. Real time communication systems would be required of the customers included in the class samples to permit their data to be obtained and used along with hundreds of others to continuously update the class load forecasts for submission to the PX.[5]

The following questions exist for Approach 2:

a. What advance notice requirements would be necessary to allow the UDC to process the transfer of a full service customer to DA status so that this information could be included within the UDC load forecasting process?

b. Would any adjustments be made for those instances in which the sum of retailer's load forecasts was

implausibly high or low, because the methods used all simultaneously erred in the same direction?

5. Summary

Developing load forecasts for submission by direct access retailers to the ISO through a scheduling coordinator or by the UDC to the Power Exchange will require changes in practices which utilities have refined under the circumstances of the integrated utility dispatching resources to meet loads while preserving a high degree of system reliability. At least two different methods of preparing a load forecast for the UDC have been discussed. Each has tradeoffs of difficulty, cost, and apparent advantages and/or disadvantages. Many of the issues facing the UDC in developing accurate load forecasts also apply to aggregation retailers who have large numbers of smaller customers for which detailed analyses of specific customers is infeasible.

It should be noted that some of the data and analysis issues discussed here for load forecasts also appear as issues in the ISO and/or scheduling coordinator settlement process, where the use of estimates versus actual load data is also under discussion.[6]

END NOTES


[1] Requests for release of customer-specific load data must be submitted to utilities and received from utilities in a timely fashion to permit analysis to support forecast development. Large numbers of such requests may be anticipated, thus utilities will have to create systems that permit such data requests to be processed satisfactorily in significantly greater volumes than have ever been requested in the past. Most requests for data will not result in hourly load data, since few customers have been metered to provide this level of specificity. Therefore, the potential retailer will have to have in place methods using assumed load profiles to convert aggregated data into hourly load schedules.

[2] The anticipated access of all electricity consumers in the United Kingdom in 1998 to competitive providers has caused the Pool to develop a series of load profiles to which each customer will be assigned for purposes of forecasting loads and allocating cost responsibility for each customer of a supplier without half-hourly metering. Source: Bill O'Reilly, United Kingdom Data Collection Service, May 17, 1996.

[3] This paper will not address the complexity of developing separate load forecasts for at least four classes of electricity consumers: (1) time-aggregated energy consumers unresponsive to real time or TOU price, (2) consumers with RTP or TOU metering and information notification capabilities, (3) customers who make explicit load bids with price caps, and (4) customers who are willing to shed load for a fee as part of the ISO's options to preserve system reliability within desired tolerances. Each of these classes of customers might exist in either direct access or full service variants, and therefore all retailers of generation services may have to make various "class" load forecasts along these lines.

[4] CPUC, D.95-12-063 categorically rejects earlier utility proposals to impose penalties or otherwise restrain DA customer return to full service customers of the UDC, p. 73. JACR of May 17, 1996 reiterates that the three UDCs retain an obligation for those customers "...who do not elect to procure their own electricity supplies.", p. 10.

[5] A similar technique may be required of aggregation retailers for developing load forecasts for their customers. While conceptually similar, the sampling issues to ensure representative data for rapidly evolving classes might be especially challenging in the initial years when direct access is permitted for smaller size customers.

[6] See M. R. Jaske, "Metering and Communication Systems in Relation to Consumer, Supplier, and System Information Requirements", Direct Access Working Group Paper, May 16, 1996._