CASE STUDIES
The ANN approach described in the previous section was
used to forecast hourly load values for a Brazilian electric
company (CEMIG - State of Minas Gerais). The test period
covers a whole year (from MarcWl993 to February/l994).
The forecasting performance is analyzed through 3 metrics:
* Mean Absolute Percentage Error (MAPE).
* Root Mean Squared Error (RMSE).
0 Theil's U (U).
approach to be tried in the fume is to use a specific network
to handle only Holidays.
Table 1 - Forecasting Error for each Month
To analyze the forecasting performance in detail, 'we
consider two particular months: March and August. Table 2
shows the MAPE, RMSE and Theil's U for each day of these
2 months and Table :3 shows the MAPE obtained for each
particular forecasting hour (forecasting with leading times
up to 24 hours).
As shown in Table 2, the MAPE is below 2% for almost
all days in both months. Also, the U coefficient is below 1
for all days, showing the ANN really extracted the more
relevant information of the load series in the training
process.
MfWE and RMSE have been commonly used to evaluate
forecasting performance; addtionally, we use a metric called
Theil'r; U, which is the rate between the RMSE of the actual
forecasting system (the ANN) and the RMSE of a "naive"
forecasting system