I am applying a deep LSTM network in order to classify time-series data from different sensors. In the field (energy) I often see the research using bidirectional LSTMs for forecasting. I don't get it how can we forecast future time steps if we need the "future" for bidirectional flow through the network. I assume creating a control algorithm using bidirectional LSTM doesn't make sense if we are interested in the current situation? For classifying past data I understand the point of bidirectional networks. Or did I get it wrong?
P.S. in my use case I get better results using unidirectional than bidirectional network, even though similar research used bi-network.