Adaptive time window convolutional neural networks concerning multiple operation modes with applications in energy efficiency predictions

Energy efficiency prediction models promote the efficient uses of energy and low consumptions of raw materials. The Convolutional neural network (CNN) is one of the most effective deep learning networks for complex process modeling. However, when applied to real industrial processes, the performance of the CNN would be restricted by the change of operating conditions, such as swings in feedstock qualities, different manufacturing strategies and variations in product specifications. A globally invariant model is unable to adapt the time-varying conditions. Therefore, we proposed a multiple operation modes adaptive time window convolutional neural network (MOM-ATWCNN). Here, a hierarchical clustering approach is suggested to determine the numbers and locations of the modes. Then, an optimal length of time window is selected to match with each mode accordingly. Lastly, the improved deep learning model is used to extract the varying features hidden in different modes. To verify the effectiveness, the proposed method is compared to several typical deep learning models by the data collected from a real industrial atmospheric and vacuum distillation process. The results show that the energy prediction accuracy of the MOM-ATWCNN is 6.5%, 2.9% and 10.2% higher than those of the traditional CNN, LSTM, BPNN, respectively. Furthermore, the proposed method exhibit its superiority regarding various performance indexes. The improvement of the algorithm is beneficial to the reduction of energy consumptions thus achieving economic goals.