Regional Heterogeneity in Energy Markets: Benchmarking Deep-Learning Models for China’s LPG Price Forecasting

Accurate liquefied petroleum gas (LPG) price forecasts are relevant to procurement and inventory planning, yet regional differences within China have received limited attention. This empirical study benchmarks 16 learned time-series models and a Naive baseline for 30-day forecasts of LPG spot prices in six Chinese regions, with drift, ETS, ARIMA, SARIMA, Ridge, LightGBM, and a futures-close reference. The dataset contains 20 daily variables and 1110 observations from 1 January 2021 to 15 January 2024. The selected regional neural configurations achieve lower OWA than ARIMA and SARIMA in five of six regions (average OWA 0.885 versus 0.984 for SARIMA). SARIMA (0.984) also improves on ARIMA (0.993). None of the twelve neural-versus-benchmark comparisons is significant after Holm correction. The selected Southern and Eastern configurations improve on Naive in both test halves, whereas the other four configurations change their relative performance across halves. These results motivate regional and temporal validation with competitive conventional baselines rather than a claim of universal neural-model superiority.