Forecasting the power consumption of a rotor spinning machine by using an adaptive squeeze and excitation convolutional neural network with imbalanced data

Abstract Forecasting the specific power consumption is crucial for cleaner production and energy saving in yarn manufacturing. A power consumption forecast can provide indicators for process parameter optimization to improve energy efficiency. In industrial scenarios, the current methods always fail to accurately forecast the abnormal values that are crucial for the prevention of unnecessary energy consumption, since the imbalanced data (the normal and abnormal power consumption sample sizes are different) misleads prediction models. Aiming to tackle the problem of imbalanced data, the root cause of abnormal power consumption is analysed to obtain the potential influencing factors of reactive power consumption in the rotor spinning process. To rebalance the prediction model, an adaptive squeeze and excitation convolutional neural network (adaptive SE-CNN) is proposed with a dual-input network architecture to discriminatively address the factors that affect the reactive power consumption (RPC) and active power consumption (APC). Then, a unilateral squeeze and excitation block is designed to automatically learn the discriminative features. It squeezes the feature information into a channel descriptor and reweights the channel feature responses with an adaptive RPC-sensitive learning rate during the training process. The experimental results based on real-world data show that the proposed approach achieves the highest prediction accuracy of 92.6252%, which indicates that 1,0331.74 k W ⋅ h of electrical energy can be saved for a rotor spinning machine over one month in the investigated workshop.