Long short-term memory neural network- decline curve analysis production forecast method for horizontal wells in tight reservoir based on sequence decomposition and reconstruction

Abstract

Oil production is a key parameter for evaluating geological potential. Conventional methods struggle to forecast nonlinear and non-stationary production sequences with a strong temporal trend due to the poor physical properties of tight reservoirs. This research proposes a forecast method that integrates a long short-term memory (LSTM) neural network with decline curve analysis (DCA). First, empirical mode decomposition (EMD) is applied to the production sequence, extracting multiple fluctuating intrinsic mode functions (IMF) related to manual construction and a residual (RES) representing reservoir energy depletion. Approximate entropy (ApEn) is then used to categorize each IMF into three groups based on sequence decomposition results, facilitating the reconstruction of the production sequence. LSTM forecasts the recombined IMF sequence, while DCA predicts the residual component. Results indicate that EMD effectively separates time trends, and both reconstructed components and residuals can be accurately predicted. Compared with stand-alone LSTM, back-propagation (BP) neural network, random-forest (RF) and convolutional-neural-network (CNN) models, the proposed method reduces production forecast errors by at least 25 %. This research incorporates signal-processing techniques and physical constraints into production forecasting. These enhancements provide a more accurate and reliable method for forecasting production in hydraulically fractured horizontal wells within tight-oil reservoirs.