I have a code in python to make energy consumption predictions using an lstm neural network in pytorch. I trained this model with 90 time series of length from one year to 1 hour to make predictions of several days (for example 7 days).

The problem is that I am not able to adjust the model so that the predictions are good or at least decent.

The inputs in addition to the history of the target variable are:
· Meteorological data: temperature, humidity, thermal sensation, wind velocity and precipitation in milimeters.
· Hour, month and weekday represented with sines and cosines.
· Holidays day (bool)
· Target variable lags of the 3 previous days.

Totally, 83 inputs columns.

The hyperparameters I am using are:
train_ratio: 0.7
seq_length: 168
batch_size: 24
epochs: 100 (Early stopping = 10)
lr: 0.001
weight_decay: 0.0001
dropout: 0.1

The architecture of the neural network is two lstm layers of 40 and 60 neurons, a linear layer of 32 neurons, a relu activation function and an output of 24.

What do you recommend to improve the predictions? They are currently very bad.