Multi-step ahead forecasting of daily urban gas load in Chengdu using a Tanimoto kernel-based NAR model and Whale optimization

Abstract

Coal-to-gas switching in urban areas plays an important role in accelerating the pace of carbon neutrality. Accurate urban gas load forecasting is beneficial in balancing the peak-valley difference and achieving high-efficiency gas utilization. This work aims to develop a new method based on Tanimoto kernel-based nonlinear autoregressive (NAR) model for dynamical modelling. The Tanimoto kernel is extended to be available for regression modelling for the first time, and of which some important properties are analyzed. Besides, a new optimization scheme based on holdout validation and Whale optimization algorithm is introduced for hyperparameter optimization. Then, the proposed Tanimoto kernel-based NAR model is applied for 5-step ahead forecasting with four regular lags 6, 9, 12, and 24 of short-(2015/1/1-2015/12/31), medium-(2014/1/1-2015/12/31), and long-(2013/1/1-2015/12/31) term daily urban gas load (UGL) in Chengdu. Results show that the proposed Tanimoto kernel-based model always produces higher precision in 80% of sub-cases than the other 11 kernel models and 8 machine learning models, with the one-step ahead forecasting mean absolute percentage error (MAPE) ranging from 2.3375% to 3.8765%, less than the other models ranging from 0.2335% to 34.5432%, and the proposed optimization scheme is efficient in improving the model’s generalization ability and robustness.