1D convolution layers can be used to develop time series forecasting models. A time series having 1 x m observations is like an image of dimension p, which has a height of a single pixel. In this case, 1D convolution can be applied as a special case of 2D convolution using a 1 x 3 filter. Additionally, the filter is moved only along the horizontal direction by strides length of 1 x 8 time units.
Let's understand how 1D convolution works. Consider the following figure that shows a time series of ten timesteps. A (1 x -1) + (2x1) + (-1 x 2) = -1 filter is moved by a stride of one time unit over the series. Thus, a 1 x 3 feature map is generated. The first element of the feature map is computed as 1 x 10. The rest of the timesteps ...
