A new model for forecasting the short-term daily demand of urban natural gas

To accurately predict the short-term urban natural gas demand is of great significance to urban gas peak-shaving and adjustment, stable supply, pipeline network optimization, etc. The present prediction models for short-term urban natural gas demand mainly include time series, regression analysis, least squares support vector machine (LS-SVM), and grey relational analysis. However, the accuracies of these models are not satisfactory except for BP neural network analysis, where its generalization is reduced by its frequent occupancy by local minimums. Comparatively, the LS-SVM prediction method based on the minimizing structural risk principle, is proved to be of higher accuracy and generalization, the most important of all, with an overfitting phenomenon rarely appeared. All involved factors that affect the short-term daily urban gas demand are considered and the three major dimensions of weather, date, and policy are determined. The fuzzy comprehensive evaluation method, the experience scoring method, and the expert scoring method were applied to deal with the qualitative data in these three factors, processed other quantitative data by the method of extreme difference transformation. A new LS-SVM-based model to predict the short-term daily demand of urban natural gas was established. Pilot tests were performed in Chengdu and demonstrated that the average absolute percentage error of prediction results was only 1.423% with this new model, the accuracy of which, compared with the ARIMA, gray correlation, BP neural network and the nonlinear regression models, was greatly improved.