From traditional to modern methods: Comparing and introducing the most powerful model for forecasting the residential natural gas demand

Abstract

Natural gas demand forecasting is of great importance for politicians and authorities, specifically in the residential sector, as a substantial percentage of residential energy consumption is related to natural gas. In order to find the best forecasting model, comparing different methods in each particular problem is crucial to evaluate the power of various models and find the most accurate model. This article has selected multi-layer perceptron and support vector machine as two neural network models and vector autoregression and multivariate generalized autoregressive conditional heteroskedastic as two econometric models to forecast monthly natural gas demand in the residential sector of Tehran province for 24 months, from the beginning of March 2019 to the end of February 2021. Six variables were used, including weather variables which strongly affect natural gas consumption and improve the accuracy of forecasts. Data for the mentioned variables have been obtained monthly from the National Iranian Gas Company and tutiempo.net from March 2003 till February 2021. The proposed models’ results have been compared using mean square error criterion (MSE), root mean square error criterion (RMSE) and Mean Absolute Percentage Error criterion (MAPE). The research findings confirm that multi-layer perceptron model had the best performance among four proposed models with the lowest measure of error.

Keywords

• Natural gas demand forecasting
• Residential sector
• Consumption
• Neural network models
• Econometric models
• Weather variables