Optimizing Production Forecasting in Dynamic Complex Industries with Deep Reinforcement Learning Approaches

Modern factories aim to boost production to increase profits and remain competitive amid intense market pressure. This goal is pursued through production planning, which not only improves process efficiency but also ensures product quality and long-term market success. Among the key tools that support this objective is production forecasting, which plays a central role in facilitating manufacturing processes. In this study, we focus on enhancing the efficiency of industrial systems by developing smart predictive models based on Deep Reinforcement Learning (DRL). Specifically, time-series data from the rotary cement kiln at the SCIMAT factory (Batna, Algeria) were used to predict clinker production. The proposed model, based on the DDQN network, outperformed both the traditional Q-learning network and the DQN. The DDQN achieved a Mean Squared Error (MSE) of 0.0010, a Root Mean Squared Error (RMSE) of 0.0316, and a coefficient of determination ([Formula: see text]) of 0.9484. In contrast, the DQN achieved an MSE of 0.0014, an RMSE of 0.0374 and an [Formula: see text] of 0.9374. The traditional Q-learning network achieved an MSE of 0.0052, an RMSE of 0.0725 and an [Formula: see text] of 0.8865. The LSTM algorithm achieved an MSE of 0.0069, an RMSE of 0.0830 and an [Formula: see text] of 0.6570. The developed model is designed to be integrated into the plant’s control system via Programmable Logic Controllers (PLCs), enabling automated operation without direct human intervention. The DQN model demonstrated strong predictive accuracy, leading to improved resource management and operational efficiency. Its application can also be extended to other industrial domains.