Mode-Aware Production Forecasting of Coalbed Methane Wells via Unsupervised Parameter-Image Clustering and Similarity-Based GAF Correction

Accurate and efficient production forecasting is critical for coalbed methane (CBM) well development and management. However, conventional engineering methods rely heavily on physical assumptions and parameter calibration, whereas direct data-driven models typically establish a global input–output mapping without explicitly distinguishing different production patterns or exploiting similarities among samples, thereby limiting their forecasting accuracy and generalizability. To address these limitations, this study proposes a structured forecasting framework integrating parameter-image representation, unsupervised clustering, lightweight classification, and similarity-based correction. Geological and development parameters are first converted into grayscale images, and spectral clustering combined with K-means++ is used to identify production-related modes. A MobileNetV2-based classifier is then trained to determine the mode category of new samples. Within the identified category, the most similar sample is retrieved, and its production sequences are encoded as Gramian Angular Field (GAF) images. Difference-image-guided posterior correction is further applied before reconstructing the corrected images into production time series. Results based on numerical simulation samples show that the proposed framework achieves high accuracy for both gas and water production forecasting, with R2 values of 0.99 and 0.99 for daily gas and daily water production, respectively. Compared with representative machine-learning and deep-learning models, the proposed method demonstrates superior accuracy and efficiency, indicating its potential for rapid CBM production forecasting.