StackedForecastLSTM

class neuralhydrology.modelzoo.stacked_forecast_lstm.StackedForecastLSTM(cfg: Config)

Bases: BaseModel

A forecasting model using stacked LSTMs for hindcast and forecast.

This is a forecasting model that uses two stacked sequential (LSTM) models to handle hindcast vs. forecast. The hindcast and forecast sequences must be the same length, and the forecast_overlap config parameter must be set to the correct overlap between these two sequences. For example, if we want to use a hindcast sequence length of 365 days to make a 7-day forecast, then seq_length and forecast_seq_length must both be set to 365, and forecast_overlap must be set to 358 (=365-7). Outputs from the hindcast LSTM are concatenated to the input sequences to the forecast LSTM. This causes a lag of length (seq_length - forecast_overlap) timesteps between the latest hindcast data and the newest forecast point, meaning that forecasts do not get information from the most recent dynamic inputs. To solve this, set the bidirectional_stacked_forecast_lstm config parameter to True, so that the hindcast LSTM runs bidirectional and therefore all outputs from the hindcast LSTM receive information from the most recent dynamic input data.

Parameters:

cfg (Config) – The run configuration.

forward(data: Dict[str, Tensor]) → Dict[str, Tensor]

Perform a forward pass on the StackedForecastLSTM model.

Parameters:

data (Dict[str, torch.Tensor]) – Dictionary, containing input features as key-value pairs.

Returns:

Model outputs and intermediate states as a dictionary.

• lstm_output_hindcast: Output sequence from the hindcast LSTM.

• lstm_output_forecast: Output sequence from the forecast LSTM.

• y_hat: Predictions over the sequence from the head layer.

Return type:

Dict[str, torch.Tensor]

Raises:

ValueError if hindcast and forecast sequences are not equal. –

module_parts = ['hindcast_embedding_net', 'forecast_embedding_net', 'forecast_lstm', 'hindcast_lstm', 'head']