Non-intrusive load monitoring (NILM) dissects smart meter data to extract individual device consumption, primarily focusing on residential users. However, energy-intensive industries also require precise load monitoring for understanding electricity usage patterns and operational states. This study introduces a novel approach employing convolutional neural networks and long short-term memory networks, enhanced with an attention mechanism and optimized using a genetic algorithm. Leveraging an industrial dataset from a Brazilian feed factory and comparing against common models, our approach demonstrates superior performance, reducing normalized disaggregation errors by at least 56.3% for six devices and increasing normalized aggregate signal errors by a minimum of 10% for three devices.