I am trying to forecast the electric load, using python SARIMAX.
I use the load history and temperature profile (as the exogenous input). Both load and temperature have the seasonality of 96 (the sample time is 15 min =seasonality is daily) as you can see for 6 days in the following figure
so I used 96 step difference to make both time series stationary and used the obtained data data to train ARIMA model (4,0,4). Finally I tested the model for the next day of the operation
the one step forecasting is very good however the dynamic forecasting seems to be awful. I know that the dynamic forecasting could not be as good as one step forecasting, however I believe that this big mismatch is not related to this point.
I really Appreciate if somebody helps me to find my possible mistake. Also the (80 steps) out of sample forecast results for the differentiated signal is shown bellow