Error predicting with n_lags>0

[Arthou]

I am having a problem making my prediction work with n_lags in NeuralProphet. Below is a snippet of the dataframe I am using. It contains an hourly data of a company's energy consumption.

ds	y
2020-09-01 01:00:00	430,5
2020-09-01 02:00:00	424,2
2020-09-01 03:00:00	429,24
...	...
2022-02-31 22:00:00	415,8
2022-02-31 23:00:00	413,28
2022-02-31 00:00:00	408,66

The code below is the code I am using to predict the energy consumption of the next month, March of 2022. The dataframe presented above is saved to the data variable.

m = NeuralProphet(growth='on',
    n_lags=3*24,
    learning_rate=0.01)

m.add_country_holidays(country_name='BR')
metrics = m.fit(df=data, freq = 'H')
future = m.make_future_dataframe(df=data, periods=31*24, n_historic_predictions = True)
forecast = m.predict(future)
fig = m.plot(forecast)
plt.show()

The code works as expected when n_lags=3*24 is commented, generating the graphic below:

But as soon as that line is present again, the code only predicts one additional period after the dataframe, independently of what is put in the periods argument in m.make_future_dataframe.

I don't know if this is intended behavior when the regressor is added, but I would appreciate any help to solve this issue.

[AchilleSoulie]

This is the normal behavior.
Using the network with the default value for n_forecast = 1 will only allow you make one-step forecast into the future.
If you want to make 324 predictions, then you should specify n_forecast = 324 (this way the network learns to make prevision n_forecast step ahead in the future).
More on this topic here: https://neuralprophet.com/html/autoregression_yosemite_temps.html#Multi-step-forecast

[ourownstory]

Thank you @AchilleSoulie for the answer!
I would just add to this that it is best to forecast as far ahead as actually needed at a time, e.g. 24h, as the model learns an explicit relationship to each forecast step.