Page 218 - 2024-Vol20-Issue2

214 | Gochhait, Sharma & Bachute

TABLE VI. TABLE VII.
TRAINING OPTIONS FOR ODISHA STATE TRAINING LAYERS FOR DELHI STATE

Layer Options Layer Parameters/Options
adam - SequenceInputLayer numFeatures, ”Name”, ”input”
GradientThreshold 1 Convolution1dLayer
InitialLearnRate BatchNormalizationLayer 11, 96, ’Padding’, ’same’
MaxEpochs 0.001 -
SequenceLength 1000 ReLULayer -
Epsilon ”longest” Convolution1dLayer
L2Regularization 1 × 10-8 BatchNormalizationLayer 20, 180, ’Padding’, ’same’
Shuffle 0.0001 -
GradientDecayFactor ”every-epoch” ReLULayer -
SquaredGradientDecayFactor 0.9 Convolution1dLayer
LearnRateDropFactor 0.999 BatchNormalizationLayer 30, 300, ’Padding’, ’same’
LearnRateDropPeriod 0.1 -
GradientThresholdMethod ReLULayer -
ResetInputNormalization 10 Convolution1dLayer
Plots ”l2norm” BatchNormalizationLayer 32, 320, ’Padding’, ’same’
-
true ReLULayer -
”training-progress” DropoutLayer 0.2
BiLSTMLayer
dataset, with an 80% training and 20% validation ratio from DropoutLayer 100, ’OutputMode’, ’sequence’
the available datasets. BiLSTMLayer 0.1
DropoutLayer
III. DEEP LEARNING MODEL FOR BiLSTMLayer 105, ’OutputMode’, ’sequence’
LONG-TERM ELECTRICITY DropoutLayer 0.2
FORECASTING FullyConnectedLayer
RegressionLayer 110, ’OutputMode’, ’sequence’
Recent advancements in deep learning algorithms, such as 0.2
convolutional neural net- works (CNNs) and recurrent neural 1
networks (RNNs), have shown remarkable efficacy in various -
domains. Among these, the Bidirectional Long Short-Term
Memory (Bi- LSTM) model has gained significant attention TABLE VIII.
as an effective RNN structure for time series prediction. Ad- TRAINING OPTIONS FOR DELHI STATE
ditionally, there is a growing trend of leveraging 1D CNNs
orcombining both 1D CNN and Bi-LSTM algorithms to en- Options Values
hance forecasting accuracy. Comparative studies have consis- GradientThreshold 1
tently demonstrated the superior performance of these models
compared to conventional statistical or machine-learning mod- InitialLearnRate 0.001
els. MaxEpochs 1000
In this section, we present a comprehensive overview of the ’longest’
theoretical foundations underlying these neural networks. We SequenceLength 1 × 10-8
delve into the mechanisms and architectural details of CNNs Epsilon 0.0001
and RNNs, with a particular focus on the Bi-LSTM model. ’every-epoch’
The aim is to provide a clear understanding of the proposed L2Regularization 0.9
models and their potential benefits for long-term electricity Shuffle 0.999
forecasting in the context of our case study of Orissa andDelhi 0.1
States. GradientDecayFactor 10
SquaredGradientDecayFactor ’l2norm’
A. CNN-BI LSTM hybrid model with deep 1D modeling true
Multivariate 1D time-series signals can be accurately pre- LearnRateDropFactor ’training-progress’
dicted by combining a hybrid model of 1D-CNN and BiL- LearnRateDropPeriod
GradientThresholdMethod
ResetInputNormalization

Plots

STM. This approach has been extensively studied and shown
promising results in various domains such as weather predic-
tion, speech recognition, stock price forecasting, and power
usage prediction. The combinationof Convolutional Neural
Networks (CNNs) and Long Short-Term Memory (LSTM)