Transformer-Based Spatio-Temporal Framework for Forecasting Natural Gas Demand in the Power Sector

Abstract

Effective forecasting of natural gas consumption in the Pakistan’s power sector is crucial for energy management and supply chain optimization. However, the non-linear and irregular behavior of natural gas consumption by the power sector makes it significantly complex, compounded by the scarcity of spatially attuned transformer models integrating regional climatic heterogeneities with temporal volatilities in import-dependent economies. To address this gap, this paper proposes the Transformer-based Spatio-Temporal Framework (TSTF), an adaptation of multiscale transformers that synergizes Chebyshev graph convolution networks (ChebyNet) to encode geospatial dependencies across urban centers (e.g., Lahore, Faisalabad, Multan) and multi-head attention mechanisms to model multiscale temporal patterns in natural gas demand in power sector. Compared to established baselines such as LSTM and GRU, the proposed model achieves superior performance, with MAE in the range of 27.99–34.44 and R2 between 0.93–0.96. The experimental results demonstrate the effectiveness of the approach in addressing climatic volatility, with practical implications for optimizing operations.