Causally interpretable multi-step time series forecasting: A new machine learning approach using simulated differential equations