Page 90 - 2023-Vol19-Issue2

86 |
Fig. 3. a block diagram of long short-term memory
ct = ft .ct-1 + it .c˜t (4)
Ot = s (Wt .Xt +Uo.ht-1 + bo) (5)
ht = Ot .tanh(ct ) (6)
The prediction strategy framework using the LSTM model
is shown in Fig .4. Three major parts make up this model:
• Preprocessing of the initial data occurs mostly at the input
layer;
• The training of the data and parameter optimization are
done using the hidden layer;
• Using the model that was developed in the hidden layer,
the output layer predicts the data.
The dataset from method was utilized to test the prediction
performance by calculating the root-mean-squared Error using
a multivariate LSTM method (RMSE). X and Y parameters
are used to access data used for training and testing. deal with
LSTM model data preprocessing. The neurons used in the
input layers are also what supervised learning uses and defines.
Adam optimizer is chosen as the gradient in the Mean Square
Error (MSE)-based model. test the accuracy and data loss of
a model after fitting it. In order to count all of the epochs and
determine the training loss, the model runs. After updating the
model’s internal states, a final model is created. Covers the
forecasting of multivariate data. For purposes of evaluation,
the RMSE is calculated as the squared difference between the
Swide & Marhoon
Fig. 4. the framework for LSTM power consumption
forecasting.[22]
actual and predicted values. increased independence, predic-
tion is more effective than bagging. It is also quicker since
every tree is trained using a subset of features. In contrast,
Bagged decision trees select to partition variables in a greedy
manner that minimizes errors. Therefore, even Bagging can
maintain a number of structural similarities, and their predic-
tions are in fact tightly connected. Predictions from the sub
models should therefore be uncorrelated or just weakly linked
in order for the ensemble of predictions from multiple models
to perform well. In order to learn the sub-trees and reduce
the connection between all sub-tree predictions, RF alters the
method. The learning algorithm is permitted to choose the
optimum split point for each variable when selecting a split
point. The RF algorithm modifies this function to assess only
a random subset of features. The following are the RF steps:
1. Create a random subset within the sample (bootstrap-
ping).
2. Choose a random set of features for each node’s best
split inside the decision tree.
3. For each subgroup, create a model decision tree.
4. Average final forecasting results and the sum of all
decision trees’ forecasts.
IV. RESULT AND DISCUSSION
A. Results of Energy Consumption Prediction Using Deep
Learning and RF Models:
This section discusses a variety of experimental findings
for all versions of models created using LSTM and RNN en-
hancement. A comparative analysis was also carried out using