China’s LNG Ex-Factory Prices by Integrating Large Language Models and Ensemble Learning: A Cutting-Edge Hybrid Approach

Abstract

China’s liquefied natural gas (LNG) ex-factory prices are affected by market-oriented reforms, supply structure changes, and diverse heterogeneous factors, presenting strong nonlinearity, nonstationarity, multiscale coupling, and high-frequency disturbances. Traditional models fail to capture such complicated dynamics, bringing difficulties to industrial capacity allocation, procurement, and inventory management. This paper proposes a CEEMDAN–LSTM–MiniMax hybrid framework combining signal decomposition, ensemble learning, and large language model semantic reasoning to realize closed-loop synergy of numerical forecasting and semantic decision-making. CEEMDAN decomposes raw price series, and LSTM models each component with validation set-based combinatorial optimization for ensemble prediction. The embedded MiniMax LLM extracts information from policies, market sentiment, and supply–demand texts to calibrate forecasts and support human–machine collaborative pricing, constructing a “numerical drive–semantic refinement–decision output” loop. Empirical results show MAE = 29.424, MAPE = 0.744%, RMSE = 38.557, and R2 = 0.991. The model outperforms single LSTM and conventional CEEMDAN ensemble models in accuracy and adaptability to extreme price volatility.