Data-driven short-term natural gas demand forecasting with machine learning techniques

Natural gas demand forecasting is one of the most crucial steps in the proper planning and operation of natural gas supply systems. The demand and supply of natural gas must be balanced at all times.Large error in forecasts of natural gas demand can cost Local Distribution Companies (LDCs) millions of dollars. In this study, techniques for accurate forecasting of natural gas demand are examined. The models are tested and validated on real data from nPower forecasting competition 2018, which consists of historical natural gas consumption and the corresponding weather forecast at 6-h intervals. The methodology presents a holistic approach that includes data pre-processing, feature engineering, feature selection, model development, and post-processing. To capture the intra-day variability in natural gas demand a block-wise approach is used to develop the forecasting models. In this approach, a separate model is developed for each block of the day. Subsequently, four different forecasting models are developed using the block-wise technique, namely, a block-wise gradient boosting model using features from sensitivity analysis (GB), a block-wise gradient boosting model using features from PCA (GB-PCA), a block-wise ANN-CG model using features from sensitivity analysis (ANN-CG) and a block-wise ANN-CG model using features from PCA (ANN-CG-PCA). Three hybrid forecasts are also developed by combining the forecasts from the four individual models. The results show that the combined models outperform the individual models, with an improvement of around 15% in terms of Mean Absolute Percentage Error (MAPE).