A Method for Predicting Reservoir Single-Well Production Based on an Improved iTransformer Deep Neural Network and Its Application

Stable production and sustained output from oilfields are critical for national energy supply and “dual carbon” goals. However, reservoir production forecasting faces significant challenges in dynamically capturing interwell spatiotemporal correlations and quantifying the informational impact of complex engineering interventions, hindering refined production management. To address these issues, this study proposes an enhanced production forecasting framework based on an improved iTransformer deep neural network for intelligent information fusion. The framework’s core innovation lies in its enhanced intrawell-area information fusion module, which employs attention mechanisms to synthesize relational data from other wells within the region, thereby strengthening the representation of interwell interference and synergistic effects. Concurrently, an intrawell-measurement enhancement module is introduced to incorporate the informational impact of key engineering measures, such as tubing hanger installation, bubble separation, and fracturing, on the target well’s daily production through time-decay functions and feature weighting mechanisms. Using daily gas production data from a real oilfield spanning 2018 to 2024, the proposed model was evaluated. Over a 96-day prediction horizon, it achieved superior performance with a mean squared error (MSE) of 0.0978, root mean squared error of 0.3127, and coefficient of determination (R2) of 0.8765. These metrics significantly outperform traditional time-series models, such as the recurrent neural network (RNN) (MSE of 0.4173, R2 of −0.0158) and the baseline iTransformer model (MSE of 0.2617, R2 of 0.7354).