Multivariate Encoder-Decoder Model approaches value - keras

Question

I am creating my first multivariate multistep encoder-decoder LSTM to forecast revenues.
As you can see, the values move towards a value and then stop at that value. The aim is to create a forecast for a longer period, but there is no deviation at all from this standard value after the first week.
What is wrong and what can I do? To me it doesn't look like it is working at all.
code:
class ModelTrainer:
def __init__(self, prediction_length=30, offset=1):
self.prediction_length = prediction_length
self.offset = offset
self._setup_values()
self.use_scaling = True
self.__prepare_data()
def _setup_values(self):
# Model configuration
self.additional_metrics = ['accuracy']
self.embedding_output_dims = 15
self.max_sequence_length = 300
self.num_distinct_words = 5000
self.verbosity_mode = 1
# DATA
self.WINDOW_LENGTH = 70 # ! SHOULD BE ADJUSTED TO THE AMOUNT OF FORECASTING DAYS
self.SAMPLING_RATE = 1
self.BATCH_SIZE = 128
# MODEL
self.DROPOUT = 0.3
self.NODES_PER_LAYER = 256
self.NUMBER_OF_LAYERS = 3
# TRAINING
self.LEARNING_RATE = 0.001
self.OPTIMIZER = Adam(learning_rate=self.LEARNING_RATE)
self.VALIDATION_SPLIT = 0.20
self.NUMBER_OF_EPOCHS = 10
self.TEST_SIZE = 0.1
self.RANDOM_STATE = 123
self.LOSS_FUNCTION = MeanSquaredError()
def __import_data(self):
self.series = DataOrganizer().df
def __prepare_data(self):
self.__import_data()
self.scaler = preprocessing.MinMaxScaler()
data_scaled = self.scaler.fit_transform(self.series)
self.features, self.target = self._create_feature_target_values_window(
data_scaled)
def _create_feature_target_values_window(self, data):
self.number_of_output_columns = 4
feature_data = data
target_data = data[:, :self.number_of_output_columns]
features, target = list(), list()
in_start = 0
for _ in range(len(data)):
in_end = in_start + self.WINDOW_LENGTH
out_end = in_end + self.prediction_length
if out_end <= len(data):
features.append(feature_data[in_start:in_end, :])
target.append(
target_data[in_end:out_end, 0:self.number_of_output_columns])
in_start += 1
return np.array(features), np.array(target)
def __create_LSTM_model(self):
num_feature_columns = self.features.shape[2]
num_output_columns = self.target.shape[2]
model = Sequential()
model.add(LSTM(self.NODES_PER_LAYER, input_shape=(
self.WINDOW_LENGTH, num_feature_columns)))
model.add(Dropout(self.DROPOUT))
model.add(RepeatVector(self.prediction_length))
model.add(LSTM(self.NODES_PER_LAYER, return_sequences=True))
model.add(Dropout(self.DROPOUT))
model.add(TimeDistributed(Dense(self.NODES_PER_LAYER)))
model.add(Dropout(self.DROPOUT))
model.add(TimeDistributed(Dense(num_output_columns)))
model.summary()
return model
def train_model(self, callbacks=[]):
model = self.__create_LSTM_model()
model.compile(loss=self.LOSS_FUNCTION,
optimizer=self.OPTIMIZER,
metrics=['accuracy', MeanAbsoluteError()]
)
model.fit(
x=self.features,
y=self.target,
epochs=self.NUMBER_OF_EPOCHS,
validation_split=self.TEST_SIZE,
shuffle=False,
callbacks=callbacks
)
self.model = model
def create_forecast(self):
prediction = self.model.predict(self.features[-1:])
# prediction = self.model.predict(self.features[-30:-29]) # Show forecast from a month old
test_X = self.features.copy()
test_X = test_X[:self.prediction_length,
:1, self.number_of_output_columns:]
test_X = test_X.reshape(
self.prediction_length, self.series.shape[1] - self.number_of_output_columns)
prediction = prediction.reshape(self.prediction_length,
self.number_of_output_columns)
inv_yhat = np.concatenate((prediction, test_X), axis=1)
inv_yhat = self.scaler.inverse_transform(inv_yhat)
prediction_df = pd.DataFrame(
inv_yhat, columns=self.scaler.feature_names_in_)
first_date = self.series.last_valid_index() + timedelta(days=1)
last_date = first_date + timedelta(days=self.prediction_length-1)
days = pd.date_range(first_date, last_date, freq='D')
prediction_df.set_index(days, inplace=True)
prediction_df = prediction_df[self.series.columns[0:4]]
Actual
Forecast:
(I know the x-axis description is incorrect. Don't worry about it)