Coal Price Forecasting Using CEEMDAN Decomposition and IFOA-Optimized LSTM Model

Zhuang Liu, Xiaotuan Li

Abstract: This study introduces a novel hybrid forecasting model for coking coal prices, integrating complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and long short-term memory (LSTM) neural networks, enhanced by an improved fruit fly optimization algorithm (IFOA). The approach begins with CEEMDAN decomposing the coking coal price sequence into intrinsic mode functions (IMFs) and a residual component, effectively mitigating non-stationarity and nonlinearity. High- and low-frequency IMFs are differentiated using a single-sample T-test, with high-frequency components consolidated to minimize noise interference. Subsequently, the IFOA algorithm optimizes LSTM hyperparameters, boosting both generalization and prediction precision. Empirical validation, leveraging the Platts price index for four major imported coking coal varieties, demonstrates that the CEEMDAN-IFOA-LSTM model significantly outperforms a broad range of benchmarks, including ANN, IFOA-LSSVR, CEEMDAN-LSTM, LSTM, BiLSTM, TCN, IFOA-LSTM, CEEMDAN-FOA-LSTM, CEEMDAN-PSO-LSTM, and CEEMDAN-GA-LSTM, achieving reduced root mean square error (RMSE) and mean absolute percentage error (MAPE). The study concludes that this model adeptly addresses the challenges of nonlinear coupling and hyperparameter optimization, offering a reliable tool for coking coal price forecasting. Future research will aim to refine the model further to adapt to diverse market conditions and enhance forecasting accuracy.

External IDs:dblp:journals/ijcisys/LiuL25