Temporal CNN / TCN

Best for: Long sequence forecasting

How it works

Applies stacks of 1-D causal convolutions so each output depends only on present and past inputs, $y_t=f(x_t,x_{t-1},\dots,x_{t-k})$, with dilation rates growing exponentially to give a large receptive field cheaply. Residual connections and causal masking preserve the time order and stabilise deep stacks, while parallel convolution makes training far faster than recurrent alternatives. A TCN matches or beats LSTMs on many sequence benchmarks while keeping a fixed, deterministic receptive field.

Common fields

Industrial IoT · signals