Multi-level decomposition-driven fusion for short-term prediction of ground well gas extraction

To address the nonlinearity, multiscale characteristics, and noise interference involved in predicting coalbed methane extraction from surface wells, a short-term methane extraction prediction model based on multi-level decomposition and a multi-model collaborative mechanism was proposed. The original signal was subjected to two-stage modal decomposition using improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and empirical wavelet transform (EWT). The decomposed components were classified into high-, medium-, and low-frequency groups using sample entropy thresholds of 2 and 1, and the high-frequency components were further refined through variational mode decomposition (VMD). An extreme learning machine (ELM), a Bayesian optimization-based bidirectional long short-term memory network (BO-BiLSTM), and a temporal convolutional network (TCN) were employed to predict the high-, medium-, and low-frequency components, respectively. Subsequently, a backpropagation neural network (BPNN) was used to integrate the outputs of the three models for ensemble prediction, thereby constructing a multi-level decomposition-driven fusion model, denoted as ELM/BO-BiLSTM/TCN-BPNN. Comparative experiments were conducted against four individual prediction models, namely ELM, BO-BiLSTM, TCN, and BPNN, to quantitatively evaluate the prediction accuracy of the proposed model. The experimental results showed that the proposed model significantly outperformed the conventional individual models in predicting short-term fluctuations, medium-term trends, and long-term stable components, thereby achieving accurate prediction across multiple temporal scales. The ELM and TCN models achieved R² values close to 1 for the high- and low-frequency components in the test set, respectively, indicating high prediction stability and accuracy. Although the BO-BiLSTM model exhibited slightly weaker generalization capability for the medium-frequency components, its R² value remained at 0.7924. Based on the comparison of prediction error metrics, the proposed multi-level decomposition-driven fusion model achieved RMSE, MAE, MSE, and MAPE values of 4.1954, 3.2554, 17.6011, and 1.0452, respectively, on the test set. Compared with the individual models, the four error metrics were reduced by up to 71.06%, 68.96%, 92.10%, and 69.64%, respectively. The proposed model achieved an R² value of 0.9366, which was significantly higher than those of the ELM, BO-BiLSTM, TCN, and BPNN models, corresponding to relative increases of 162.50%, 285.91%, 375.91%, and 305.45%, respectively. These results demonstrate that the proposed model provides superior prediction accuracy, stability, and generalization capability compared with conventional individual models, thereby supporting accurate multi-timescale forecasting for field coalbed methane extraction through surface wells.