Seite - 72 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

Table 1 Models for forecasting evaluation Model 1 Naive 7tt CC Model 2 Holt Winters Multiplicative skTkt ckTbTaC where a, b and c are the estimated recursive coefficients Model 3 Static Model with meteorological variables From equation 1 Model 4 Dynamic Model with meteorological variables From equation 2 Model 5 Static Model with meteorological variables and HPDI From equation 3 Model 6 Dynamic Model with meteorological variables and HPDI From equation 4 Model 7 Static Model with meteorological variables and daily airport arrivals From equation 3 Model 8 Dynamic Model with meteorological variables and daily airport arrivals From equation 4 Comparison of the different sets of forecasts was undertaken using the mean absolute percentage error (MAPE) summary measure, which is used extensively in the electricity demand forecasting literature. It is important to highlight how the performance of the exercise was assessed. On the one hand it was assessed by the whole year while on the other hand because of the special interest for the electric system in providing good accuracy forecasting during the peak seasons the data base of errors was split into three parts: the high season, which comprises June, July, August and September; the low season, which covers January, February, November and December; and the mid-season, consisting of March, April, May and October. 4. Results and forecast performance Daily forecasts were used to set up the weekly network outage plan. They were computed by the middle of the week, usually on Wednesday morning, with information up to Tuesday, for the seven-day period beginning the following Saturday. The relevant lead times go from 4 to 10 days ahead, although they are some minor modifications. For example, the origin of the forecast changes when a public holiday falls on Wednesday. I ignore this for the sake of simplicity, and act as if there is a one to one relationship between the day of the week and the lead time. Thus, table 2 reports MAPE for the different models used in this study. Bold figures indicate which model attains the lowest MAPE. M.BakhatandJ.RosselloNadal / ImprovingDailyElectricityLoadsForecasting72