Enhancing Sugar Factory Efficiency with Deep Learning-Driven Control Strategies

Abstract

Sugarcane crushing systems are crucial in sugar production because the initial juice extraction step significantly impacts overall efficiency and productivity. Conventional control strategies often fall short in addressing the dynamic and nonlinear nature of these processes, leading to high energy consumption and suboptimal juice yields under variable operating conditions. This paper proposes an LSTM-based predictive control framework to enhance sugarcane crushing efficiency by integrating mathematical system modelling, deep learning, and advanced optimization principles. A synthetic dataset representing key nonlinearities (feed rate, roller speed, and extraction pressure) is used to train the LSTM, capturing critical relationships between inputs and outputs—juice yield and energy consumption. Model Predictive Control (MPC) further refines the control inputs in real-time, ensuring adaptability to fluctuating feedstock quality and equipment wear. Simulation results demonstrate a systematic reduction in training loss, reasonably accurate yield and energy predictions, and the potential for significant performance improvements over traditional controllers. While the synthetic nature of the data underscores the need for real plant validation, this work provides a foundation for developing robust, adaptive, and data-driven strategies capable of transforming sugarcane crushing operations.