Residual-based ARIMA_CNN_LSTM modeling for borehole stress time-series forecasting

Abstract

This research establishes are residual-based ARIMA–CNN–LSTM model for predicting borehole stress in non-stationary, nonlinear, and noise-affected monitoring situations. The ARIMA (2,1,5) initially models the foreseeable linear and autocorrelated part. Subsequently, a CNN–LSTM network captures short-term patterns and long-term dependencies left in the residual. The stored dataset consists of 2,975 data points sampled at intervals of 15 minutes. The final 595 data points result in 594 matched one-step-ahead forecast-observation pairs. In this reserved block, the combined model attains MAE = 12.34 SU and RMSE = 15.00 SU. When compared with 51.72 and 60.26 SU of the independent ARIMA, it leads to reductions of 76.14 % and 75.11 %. In comparison to the stored signal span, NMAE and NRMSE are 0.86 % and 1.0 5%.Since the acquisition export fails to maintain a verified channel-to-MPa conversion, observations, forecasts, residuals, MAE, and RMSE are constantly reported in software units (SU). An analysis of post hoc sensitivity encompassing filter quantity, recurrent depth, and learning rate reveals only slight changes in validation error, advocating for an accuracy–complexity interpretation instead of an assertion of universal architectural optimality. The outcomes show that residual-driven statistical-deep learning is able to enhance borehole-stress forecasting, whereas multi-location validation, calibrated collection metadata, and systematic examinations of sampling and data absence are still required.