CNN LSTM - Avesta-Ahmadi/Univariate-Time-Series-Forecasting GitHub Wiki

CNN-LSTM is composed of conv1d layers and maxpooling on input data followed by stacked LSTM layers. The conv1D layer attempts to learn and extract the features according to the filters. The input to the CNN has the following shape: [samples, subsequences, timesteps, features].

Number of features is determined by number of features in time series. if it is univariate, n_features is equal to 1. If each input has 10 timesteps in it for instance, it could be 1 sequence or we can divide it in 2 subsequences with 5 timesteps.

The schematic of a CNN-LSTM network is as follows:

Source: https://towardsdatascience.com/get-started-with-using-cnn-lstm-for-forecasting-6f0f4dde5826