Research on Q5500thermal coal price prediction at Qinhuangdao Port based on AGQPSO-CNN-LSTM

LI Hui; YANG Can

Abstract：Under the deepening national"Dual Carbon"strategy,coal price serves as a"dynamic balancing pivot"in energy transition,whose stability directly impacts the feasibility of strategic goals.To address the nonlinearity,high volatility of coal prices,and the insufficient accuracy of traditional prediction models,this study proposes a hybrid prediction model integrating Adaptive Gaussian Quantum Particle Swarm Optimization(AGQPSO)with Convolutional Neural Network-Long Short-Term Memory(CNN-LSTM).By leveraging AGQPSO to globally optimize CNN-LSTM hyperparameters,this approach resolves the curse of dimensionality in traditional grid search methods.It combines CNN's spatial feature extraction capability with LSTM's temporal dependency modeling to establish a spatiotemporal deep-learning framework.Using the Q5500 thermal coal price at Qinhuangdao Port as an empirical case,time series-derived features such as first-order difference and 5-day moving average/standard deviation are introduced.Through a two-stage feature optimization strategy,the five most relevant features are ultimately selected for prediction.Experimental results indicate that the AGQPSO-CNN-LSTM hybrid model achieves excellent performance in short-term coal price prediction,with an R2 of 94.90％,MSE of 1.837 9,and MAPE of only 0.162 8％.It significantly outperforms standalone CNN and LSTM models and can serve as a high-precision and robust tool for short-term coal price forecasting.