Abstract
<title>Abstract</title> <p>Accurate production forecasting in offshore reservoirs is essential for short-term reservoir management and production planning. However, forecasting remains challenging because production responses are affected by complex transient behavior, interwell interference, and operational interventions. In particular, shutting in a producing well alters the reservoir pressure field and redistributes fluid flow, potentially increasing production rates in neighboring active wells. Under these conditions, conventional data-driven forecasting models may exhibit reduced accuracy and substantially underestimate production rates. Despite the importance of these transient responses, previous studies have largely focused on model architecture and input selection, while the influence of the training loss function and the unequal consequences of underprediction and overprediction have received limited attention. To address this gap, this study systematically compares thirteen symmetric and asymmetric loss functions for short-term production forecasting across four recurrent and convolutional deep learning architectures. The evaluation uses a complex offshore reservoir benchmark that incorporates various combinations of production and injection data. For oil production rates, asymmetric loss functions reduce forecasting errors by approximately 50% compared with conventional losses. For the more challenging water-rate forecasting problem, the best asymmetric formulation reduces error by approximately 5% relative to the corresponding baseline. Asymmetric losses also reduce the occurrence and magnitude of underprediction during positive production transients while maintaining stable forecasts under rapidly changing operating conditions. These findings demonstrate that accounting for directional forecast errors through the training objective can improve short-term production forecasting under interwell interactions and operational interventions.</p>