Rob J Hyndman

Talk to be given at the Uni­ver­sity of Mel­bourne at 1pm, Tues­day 11 Octo­ber 2011.
Loc­a­tion: Room 213, Richard Berry Build­ing, Uni­ver­sity of Melbourne.

Abstract:

Elec­tri­city demand fore­cast­ing plays an import­ant role in short-​​term load alloc­a­tion and long-​​term plan­ning for future gen­er­a­tion facil­it­ies and trans­mis­sion aug­ment­a­tion. Plan­ners must adopt a prob­ab­il­istic view of poten­tial peak demand levels, there­fore dens­ity fore­casts (provid­ing estim­ates of the full prob­ab­il­ity dis­tri­bu­tions of the pos­sible future val­ues of the demand) are more help­ful than point fore­casts, and are neces­sary for util­it­ies to eval­u­ate and hedge the fin­an­cial risk accrued by demand vari­ab­il­ity and fore­cast­ing uncertainty.

Elec­tri­city demand in a given sea­son is sub­ject to a range of uncer­tain­ties, includ­ing under­ly­ing pop­u­la­tion growth, chan­ging tech­no­logy, eco­nomic con­di­tions, pre­vail­ing weather con­di­tions (and the tim­ing of those con­di­tions), as well as the gen­eral ran­dom­ness inher­ent in indi­vidual usage. It is also sub­ject to some known cal­en­dar effects due to the time of day, day of week, time of year, and pub­lic holidays.

I will describe a com­pre­hens­ive fore­cast­ing solu­tion designed to take all the avail­able inform­a­tion into account, and to provide fore­cast dis­tri­bu­tions from a few hours ahead to a few dec­ades ahead. We use semi-​​parametric addit­ive mod­els to estim­ate the rela­tion­ships between demand and the cov­ari­ates, includ­ing tem­per­at­ures, cal­en­dar effects and some demo­graphic and eco­nomic vari­ables. Then we fore­cast the demand dis­tri­bu­tions using a mix­ture of tem­per­at­ure sim­u­la­tion, assumed future eco­nomic scen­arios, and resid­ual boot­strap­ping. The tem­per­at­ure sim­u­la­tion is imple­men­ted through a new sea­sonal boot­strap­ping method with vari­able blocks.

The model is being used by the state energy mar­ket oper­at­ors and some elec­tri­city sup­ply com­pan­ies to fore­cast the prob­ab­il­ity dis­tri­bu­tion of elec­tri­city demand in vari­ous regions of Aus­tralia. It also under­pinned the Vic­torian Vis­ion 2030 energy strategy.

We eval­u­ate the per­form­ance of the model by com­par­ing the fore­cast dis­tri­bu­tions with the actual demand in some pre­vi­ous years. An import­ant aspect of these eval­u­ations is to find a way to meas­ure the accur­acy of dens­ity fore­casts and extreme quantile forecasts.

Down­load slides