Time series outlier detection for short-term electricity load demand forecasting

Forecasting of working-days’ electricity demand is vital for short-term planning. However, demand variations due to outliers can reduce the accuracy of forecasts. Therefore, a time series data cleaning technique is proposed to remove these disturbances of electricity data. First, holidays’ and bridging holidays’ data are replaced by Moving Average. The k-sliding window filtering band is proposed to detect the time series outliers and replace by forecasted regular load demand using Moving Average. Data from the Electricity Generating Authority of Thailand (EGAT) and a Neural Network (NN) model with six inputs and one output are used to demonstrate the performance of time window data cleaning process. The sample dataset contains data from 1stMay 2012 to 31stMay 2013 where May 2013 is used for testing. The Time-Window based data cleaning technique increases the performance of forecasting outcomes by 11.60% for non-holidays. Results from the proposed technique are compared with the results from the robust version of locally weighted smoothing (r-LOESS) and identified that the proposed technique is superior for taking results for non-holidays.