Naive forecasting of household natural gas consumption with sliding window approach

Abstract

Household consumption has a significant importance for natural gas wholesale companies. These companies make one-day-ahead forecasting daily. However, there are penalties depending on the error of the estimates. These penalties increase exponentially depending on the error rate. Several studies have been done to develop mathematical models to forecast natural gas consumption and minimize the error rate. However, before mathematical model predic- tions, a previous step, data preparation, is also important. The data must be prepared correctly before the mathematical model. At this point, prior to the mathematical model, selecting the appropriate data set size has a vital role. In this study, one-day-ahead household natural gas consumption is forecasted for different data sizes. Forecasts have been made for the year 2012. For removing insignificant variables, multiple linear regression (MLR) is applied to all data. In this research, 2 particular scenarios are applied for forecasting. In the first scenario, 2 different data set models are prepared. These sets consist of the data collected 6 weeks before the forecasted day. Daily outcomes are added to the data set and the set is applied in a model called Model A. The other model is depicted based on a sliding window idea having 6 weeks of fixed data size with dynamic data inside (Model W6). For the two models, MLR is applied and error rates are compared. Here, Model A has 7 times higher mean absolute percent error (MAPE) than Model W6. In scenario 2, 6 models are studied and compared for the sliding window approach. The models are named according to the weeks involved (e.g., Model W1, Model W6). MAPEs for Model W3, Model W4, Model W5, and Model W6 are obtained as 11.8%, 6.8%, 7.2%, and 8.1%, respectively. The lowest preday error occurs in the 4-week data model with sliding window approach.