Abstract: The learning capability of such convolutional network models as temporal convolutional network (TCN) is limited to the scale of receptive field, and sometimes it is difficult to mine data from seasonal or long-term information, such as weather data for deep convolutional networks. In order to address this problem, a novel multi-scale and bilinear convolutional neural network was proposed, which was composed of the TCNs layer and bilinear layer. The TCNs layer contained two temporal convolutional networks, each of which extracted useful features independently from the historical data. All TCNs had the same parameters expect for the kernel size, thereby further mining the underlying relevant information in data. The output of the TCNs layer was employed as the input of the bilinear layer, and after normalization, the final output could be eventually calculated, which enhanced feature representation capabilities of the model. The comparisons with five methods on the public datasets of weather forecast show that the proposed method can effectively improve the accuracy and performance in terms of the long-time data much better than the TCN.

Key words: temporal convolutional network, multi-scale convolutional, bilinear fusion, data mining; weather forecast