Recurrent neural networks (RNNs) have been predominately employed to deal with industrial process operation modeling problems that are hard to be described by first-principles approaches. However, the substantial amount of iterative calculations in RNNs dramatically increases the training difficulty and reduces the speed of the network convergence. Thus, RNNs are an increasingly ineffective method when handling a large amount of historical operation data with a long operating horizon. In contrast, convolutional neural networks (CNNs) are known for fast computation but have rarely been utilized to account for time series information. To adopt the benefit of CNNs to handle the industrial time series trend prediction, in this study, we propose an adapted receptive field temporal convolutional network with bar-shaped structures (BS-ARFTCN). Based on the temporal convolutional network (TCN), a reshaped CNN structure, BS-ARFTCN, replaces the convolution windows with the translation-only sliding bar-shaped windows tailored to the specific dimensionality and functionality. As a result, this network allows the states to be predicted by its historical information, specified by the adapted receptive field in the TCN, in a computationally efficient manner. In addition, the Tennessee Eastman Process is used to validate the performance of the BS-ARFTCN algorithm, in which the result is compared with that of the long short-term memory network and the RNN. The BS-ARFTCN algorithm is demonstrated to provide a more accurate prediction with less computational resources and training time.