Weird prediction after MinMax Scaler

Question

I am trying to make forecasting with LSTM but when I do MinMax Normalization my prediction is being terrible. When I check autocorrelation, data looks stationary before and after MinMax normalization, but still result is not acceptable. (i tried almost every activation functions and many lstm layers too)

Autocorellation screenshot

Prediction

My LSTM archtiecture :

EPOCHS = 1000

steps = int( np.ceil(x_train_multi.shape[0] / batch_size) )
val_steps=int( np.ceil(x_val_multi.shape[0] / batch_size) )


multi_step_model = tf.keras.models.Sequential()
multi_step_model.add(LSTM(128,activation="relu",return_sequences=False,input_shape=x_train_multi.shape[-2:]))
multi_step_model.add(Dropout(0.4))
multi_step_model.add(tf.keras.layers.Dense(future_target)) # for 72 outputs
multi_step_model.compile(optimizer=tf.keras.optimizers.Adam(), loss='mae')

multi_step_model.summary()
print(train_data_multi)

multi_step_history = multi_step_model.fit(train_data_multi, epochs=EPOCHS,
                                          steps_per_epoch=steps,
                                          validation_data=val_data_multi,
                                          validation_steps=val_steps)