The approach

You have trained a Bidirectional LSTM model with almost 2 and a half years' worth of data (October 2015 – March 2018). You reserved the last 13 weeks for testing (July – September 2018) and the prior 13 weeks to that for validation (April – June 2018). This made sense because the combined testing and validation datasets align well with the highway expansion project's expected conditions (May – October). You wondered about using other splitting schemes that leveraged only the data representative of these conditions, but you didn't want to reduce the training data so drastically, and maybe they might need it for winter predictions after all. A look-back window defines how much past data a time series model has access to. You chose 672 hours (4 weeks) as the look-back window size because as the model moves forward, it can learn daily and weekly seasonality, as well as some trends and patterns that can only be observed across several...