International Journal of Power Electronics and Drive Systems (IJPEDS)

Vol. 16, No. 1, March 2025, pp. 633~641

ISSN: 2088-8694, DOI: 10.11591/ijpeds.v16.i1.pp633-641  633

Global solar energy estimation using improved greedy based

genetic algorithm with deep convolutional neural network

Tintu Beena Devaraj1, Charu Navaneeth Raveendran Kottoor2

1
Department of Electrical and Electronics Engineering, Kerala Technological University, Kerala, India
2
Department of Applied Technology, Institute of Applied Technology, Al Ain, United Arab Emirates

Article Info ABSTRACT

Article history: Demand for solar energy increases and it is required to manage the supply of
energy effectively. Accurate detection on patterns of energy consumed assist
Received Jan 6, 2024 in taking appropriate decisions on generating energy. Even though many
Revised Oct 9, 2024 traditional techniques have predicted the consumption rate, still improvement
Accepted Nov 28, 2024 is needed in prediction accuracy. The pre-processing is performed initially for
handling missing values. The feature selection is accomplished using
improved greedy based genetic algorithm (GGA) to extract best features to
Keywords: enhance the performance of the model. Output from feature-selection is
passed as input to the classification phase using proposed deep convolutional
Classification neural network (CNN) in which future solar energy patterns are classified and
Deep convolutional neural predicted timely basis power consumption and it optimize the model by
networks minimize the error. The prediction accuracy is estimated through evaluation
Deep learning metrics such as mean square error (MSE), mean absolute percentage error
Improved greedy based genetic (MAPE) and root mean square error (RMSE) 0.423, 0.652, and 0.215,
algorithm respectively. The outcomes achieved in terms of accuracy at 99.75, precision
Machine learning at 99.28, sensitivity, and recall at 100 revealed the efficiency of the proposed
Prediction classification model. As a result, the proposed future prediction of solar
energy was considered efficient since it achieved reduced error values than
other prediction algorithms. It assists in maintaining stability in solar-energy
based systems.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Tintu Beena Devaraj
Department of Electrical and Electronics Engineering, Kerala Technological University
Kerala, India
Email: [email protected]

1. INTRODUCTION
Managing adequate generation of solar energy based on demand patterns makes smoother flow of
electricity during peak requests for solar energy. It is accomplished through efficient prediction of energy flow
in peak operating hours in the energy-based systems. In this manner, the suggested research [1] has focused on
forecasting of electrical energy consumption generated by the renewable energy sources based on hourly basis.
Prediction has been exploited through machine learning (ML) algorithm which has optimally allocated the
operational hours of water pumps with the intention of reducing the predicted peak. Regular monitoring and
control on devices have reduced the weekly and daily deviations on electrical consumption greater than 15%.
Rate of load transferred, total energy produced through renewable sources, and accuracy in every day has
determined the potential outcome of the prediction model. Similarly, recommended research [2] have used an
open-dataset for measuring of energy usage, production of solar energy along with building of air-leakage data
from commercial buildings. Solar energy collected information from 6 years of time have been measured as 1
hour through micro-inverters based on monthly and daily predictions. It has utilized the specific dataset for

Journal homepage: http://ijpeds.iaescore.com

634  ISSN: 2088-8694

prediction and estimation of operational outcomes. Mismatches occurred between the drained water heat-
recovery devices and rate of consumption has been considered an important factor which have been resolved
through considered probabilistic method [3] of estimating the monthly basis consumption strategy. It has
observed from 113-residences for the prediction of consumption patterns. It has performed assessment of the
system in order to determine the uncontrolled supply, especially from solar energy.
Earlier predictions on energy usage through analyzing the solar radiation and demand patterns assist in
managing renewable energy effectively [4], [5]. To this purpose, the suggested study [6] has utilized long short-
term memory (LSTM) based neural network (NN) which has been enhanced through attention and
decomposition of input data. Selective concentration to the input patterns has been given for the analysis of the
energy consumed daily. Renewable energy sources like fuel cells, solar -panels for effective storing of energy
and ground-sourced heater pumps have been used in extensive applications for preventing greenhouse gas
emissions and providing sustainability [7], [8]. Investigations on suggested research [9] have analyzed the
impacts of renewable energy on the constructional building along with the flow of electricity in order to satisfy
the total amount of demand in energy. The optimal source of energy required for every block has been estimated
through linear programming. Energy loads have determined variation in the flow during the day time and the
maximum level of heater demands occur in winter morning has also been analyzed. Though, several data-driven
methods have been utilized for building energy prediction models, enhancement in robustness, generalization
and accuracy has been found to be challenging, and in the suggested research [10] support vector-based
regression technique has been used in which the higher non-linearity has been approximated by the linearity
with the multiple distortions in the high dimensional data mapped with the vector. Larger office building [11]
with hourly cooling on summer has been recorded for analysis. The nonlinear and volatile nature of power
demand and supply has occurred due to fluctuating weather conditions such as temperature and solar radiation.
In considered analysis [12], optimal extreme learning machine (ELM) has been employed for forecasting the
volatile and non-linear solar-energy prediction with variable weather conditions. Performance on the ELM
approach has been explored through relevant attributes such as hidden layers, biases, and weights. teaching and
learning-based optimization (TLBO) have been used for handling the computational complexities. Integration
of ELM with TLBO has forecasted the solar-energy generation based on hourly, daily, and monthly. The
performance of the model has been assessed by evaluating error metrics such as MSE, MAE, and MAPE [13].
Although traditional studies have delivered many methods and techniques in prediction analysis of
energy consumption to provide the energy source based on the demand, the minimum error rate in the
prediction using classification still requires improvement for accurate prediction of solar energy consumption
patterns [4], [14]. Moreover, the consumed energy prediction alone has not assisted in minimizing the
overloading condition but is also required for maintaining adequate energy sources to supply for the demand
based on hourly, daily, and weekly predictions.
To overcome the above-mentioned issues, ML and deep learning (DL) based algorithm is used. The
feature selection uses an improved greedy-based genetic algorithm (GGA) to choose the best dataset features
for analysis. DL-based deep convolutional neural network (CNN) for prediction algorithm is used for
classification process and solar energy patterns prediction with effective pre-processing of input data is used
in present research for enhancing the accuracy. The main intention of the proposed research is to enhance the
prediction efficiency of energy consumption based on solar energy sources. The utilization of advanced
algorithms such as the improved GGA for feature selection and deep CNN for classification indicates a step
towards more accurate and efficient energy management systems and addresses the challenges of conventional
studies in gaining efficient future prediction of solar energy consumed. Depending on the analysis made from
different approaches, the motivation of the present research is being framed as follows:
- To perform prediction of the solar energy consumption with the input data collected from the global solar
energy dataset observed from New Delhi and Kolkata and perform pre-processing with categorical
encoding.
- To execute feature selection using an improved greedy-based genetic algorithm (GGA) for selecting the
best features required for the future prediction of solar energy consumption patterns and also for minimizing
the complexity by reducing dimensions.
- To accomplish the classification task with a deep convolutional neural network (CNN) algorithm to classify
the accurate electric power consumed in a timely basis with minimized error.
- To perform predictions on hourly, daily, and monthly time duration and evaluate the loss obtained in the
prediction process with the performance metrics like accuracy, root mean square error (RMSE), mean
absolute error (MAE), and mean square error (MSE) to determine the efficacy of the model.

Int J Pow Elec & Dri Syst, Vol. 16, No. 1, March 2025: 633-641
Int J Pow Elec & Dri Syst ISSN: 2088-8694  635

2. LITERATURE REVIEW
The succeeding section discusses on several existing methodologies used in the process of predicting
future energy patterns with the collected energy information since accurate forecast on solar-energy is required
for maintaining reliable operation on energy-systems. For this purpose, suggested research [15] has utilized
hybrid technique of sequential one-step ahead-models which accumulated 144 hour supply. Different inputs
have been covered over 6 days of forecasting area in which every horizon has been trained for every one hour.
It has utilized XGBoost and CatBoost methods for boosting the accuracy in prediction. Outcomes have
highlighted the models with previous month, current month of solar energy predictions to forecast energy
consumption patterns of 1-month ahead. Renewable energy load has been forecasted with the mixture of
various load patterns such as institutional, commercial, energy systems, and residential areas for reliable supply
of energy. The operational behavior of the solar energy has been administered in considered research [16].
Using daylight as the fundamental source in residents has served to administer the daily needs of lights and
seasonal variations to accompany the occupants about the rate of consumption.
Climate based daylight modelling (CBDM) in recommended research [17] have performed
evaluations on the hourly basis to assess the extent of daylight requirement. In order to improve the
performance of systems, forecasting the future demand has been found to be a significant task. For predicting,
considered research [18] have utilized deep neural networks for prediction of one day ahead of global-
horizontal irradiation. It has performed predictions with the daily recorded data and has analyzed on parameters
affecting accuracy. Forecasting performance has been assessed with error metrics which have exhibited
efficacy of the prediction model. Aggregation on demand and supply forecast has helped in controlling the
generation of solar energy and assisted in effective storage. It has been found to be mandatory for avoiding
both oversized and under-sized of energy systems. The result in [19] has carried out five year of energy
consumption data and processed with MATLAB for calculation of demand for satisfying future energy needs.
Every day capacity has been realized and has predicted that demand for solar energy has increased in five years
of time. Demand growth has been estimated over the period of time and the productive power has been coupled
with demand on 3 hours for morning power supply and two hours of mid-day and 4 hours of evening supply.
Solar-energy prediction has been considered as significant for development of solar power plants through its
capability in meeting the demands of the user. Considered research [20] has presented a methodology in
predicting solar-energy based on ML and DL techniques. It has evaluated on short term and real time
predictions but it has ensured optimized-prediction on energy. Through analysis on dataset for timely
consumption, Pearson-correlation has been deployed for identification of relevancy in rate of consumption at
particular peak demand levels.
Production of power from solar radiation has been predicted with the integration of ML technique
with the regression based methodology of forecasting 1 hour ahead of solar power [21]. Even though annual
and seasonal forecasting has been performed with different ML algorithms, the regression on daily basis has
classified the original dataset into training and testing of subset for enhancing the prediction accuracy. In order
to reduce carbon-emissions and demand level energy supervision, the ability of predictions in different patterns
of load has been identified which also significant. It has analyzed solar-energy utilizing areas for figuring out
the alterations in the energy demand and impacting factors for changing energy utilities. It also has estimated
energy requirement in medium- and long-term utilities. Energy based systems have required the forecasting
information of power consumption in different patterns and moreover, the accurate prediction of monthly usage
assist in the prediction of medium to long term demand and contributes considerably in the process of dispatch
and energy management in solar energy systems. Consequently, larger datasets have been required for the
determination of accurate prediction by the methodology and has been affected in traditional approaches with
the scarce data. Due to insufficiency in data, complications in forecasting have increased. By analyzing all
drawbacks in existing studies, proposed research is being designed to provide effective prediction accuracy.
The review conducted on traditional research have utilized several artificial intelligence-based
prediction methods for prediction of energy consumption with information collected on daily, hourly and
monthly. The common limitations with respect to validation on prediction have been analyzed from the
traditional literatures and are given as follows:
- Multi-step CNN based solar energy prediction mechanism has extracted consumption of energy data but
needs to be enriched with larger datasets which can provide effective outcomes in the demand based
prediction analysis [21], [22].
- The prediction has been performed with various fluctuations in utilities for everyday classification with ML
methods on forecasting 1 hour ahead of global solar radiation can enhance the prediction accuracy with
efficient methods [19], [23].

Global solar energy estimation using improved greedy based genetic algorithm … (Tintu Beena Devaraj)
636  ISSN: 2088-8694

3. PROPOSED METHODOLOGY
The proposed energy management approach categorizes the consumption of energy on every month
through patterns of demand and supply based on features associated with solar energy sources capable of
achieving greater prediction efficiency. Existing researchers have predicted the consumption of energy but are
inadequate with factors of accuracy and loss in prediction based on the perspective of energy constraints. To
this purpose, proposed research is motivated to provide techniques involving data pre-processing of global
solar energy dataset, feature selection using improved combination of genetic algorithm (GA) and the greedy-
sequential process and classification using deep CNN algorithm for solving energy consumption and
distribution efficiency, which extends battery lifetime. However, providing solutions in energy management
alone cannot afford the complete level of efficiency of the model. Hence, the present research intends to focus
on adopting future energy prediction mechanism for stable supply of solar energy to systems that highlights
the forward-thinking nature of the research. Figure 1 provides a clear framework of the proposed methodology,
showcasing a structured approach to predictive energy management. The overall architectural diagram of the
proposed prediction model is illustrated in Figure 1.
As depicted in Figure 1, the proposed prediction methodology represents the mechanism which is
divided into three stages of implementation such as pre-processing, feature selection, and classification. First,
input solar energy features are taken from the global solar energy dataset the initial phase of processing input
features with checking of missing values and label encoding to minimize the model's complexity and to prevent
the biased model development. With the processed data, selecting features is accomplished using improved
GGA to reduce computation complexity in the accurate prediction of solar energy. Such selected features are
consequently passed on into the third phase of classification, in which the deep CNN algorithm is used in
classifying the consumption pattern in different months. The advantage of extracting features is that it enhances
the classification accuracy and estimates the minimum loss acquired through the prediction model. For
elucidating the efficiency of the present prediction method, validation is being accomplished with various
evaluation metrics such as MSE, accuracy, MAE, and RMSE internally, and classification efficiency is being
compared with other algorithms for determining the effectiveness of the present research. Prediction outcomes
were analyzed at hourly, daily, and monthly intervals to evaluate the efficacy of the system.

Figure 1. Overall architectural diagram of the proposed model

3.1. Feature selection by improved GGA

The proposed work utilizes improved greedy genetic algorithm (GGA) algorithm was used for feature
selection to identify the most relevant temporal and complex features for analysis and selecting the best features
by integrating genetic algorithm (GA) with the greedy-based strategy. The procedure of predicting out best
feature choices at each stage is the importance of the greedy approach. Such best energy features from every
data sequence are being grouped for predicting the energy demand fluctuations to distinguish normal and peak

Int J Pow Elec & Dri Syst, Vol. 16, No. 1, March 2025: 633-641
Int J Pow Elec & Dri Syst ISSN: 2088-8694  637

flow patterns. The combination of GA and the greedy-sequential process is being utilized to solve the
difficulties with the maximum and extreme search space exploration. The greedy approach is induced as the
operator into the genetic algorithm in accelerating the searching capabilities and enhances chromosome
fitness appropriately. Crossover operation performed between the individual results in the exploitation between
two given parents. Gene in the chromosome is being modified by organizing the diverse nature of individuals
among the population. The working nature of improved GGA is presented in Algorithm 1.

Algorithm 1. Improved greedy based genetic algorithm

Begin
Generate the initial population POPii (0);
Estimate POPii (0);
Repeat
Select parents;
Generate new chromosomes using crossover;
Applied mutation on the new chromosomes;
Applied greedy sequential function;
Estimate POPii (ti);
Until (Terminating condition is reached);
End;

As illustrated in Algorithm 1, the operational process starts with the population generation by random
distribution for increasing diversity. Accordingly, the population is encompassed with multiple solutions
representing the chromosomes of individuals. Each chromosome has a finite variable group simulating genes.
Distribution of solutions around search space aggregates diversity and improves the prediction of promising
segments. The selection of the fittest individuals leads to the genes contributing to the creation of the next
generation. Until the algorithm reaches the point of criterion, it improves the population with three significant
operators. Consequently, the best solution from the finally retrieved population counts returns the best global
optimal path for the problem. Controlling attributes in the method are being changed depending on the detecting
performance. Along with that, crossover operation is also significant for attaining efficient outcomes. For
identification, chromatic numbers are being assigned. Improved GGA begins with the initial population of
POPii (0) with a feasible group of energy attributes. The objective function is represented as OFi through which
the population group is obtained through the effective selection of features that predict the degradation in
supply of solar energy through changes predicted from the influencing weather data. The function of the initial
population is formulated by generating various chromosomes with different parameters in a random manner.
The selection method is used in which two significant individuals are selected randomly from the
group population and picks out the needed features as the parent. With the assistance of acrossover operator,
two various parents with two-children remain to be generated with the genetic element of parents. Through
the GA method, three-point operators are effectively used in selecting optimal features. Uniformity in mutation
assists in introducing and preserving diversity in genetic population and it is being applied with smaller-
probability. Along with that, GA utilizes the greedy process in which verification of constraints on the problem
for each child node is not needed for all genes. A distinguished attribute acquired from a neighbor indicates
that it has been affected in certain cases. However, it must be among the utilized attributes in chromosomes,
and modification is being performed if no such occurrences are identified. The hybridization of greedy with
GA selects the best energy consumption parameters required for predicting the flow of energy. Outcome
obtained are passed on to the classification process to predict the average demand and supply patterns of energy.

3.2. Classification by deep CNN

DCNN falls under deep neural networks (DNN), used mainly for identifying patterns in input energy
data. The classification process is undertaken using deep CNN technique by utilizing a three-dimensional
neural pattern analysis. In order to maintain stability in power supply to solar energy-based systems, disruptions
in supply due to sudden peak flow are required to be predicted and accomplished with the deep CNN algorithm.
Through fluctuations in energy supply or due to other factors, performance is affected and leads to an unstable
condition. The proposed algorithm effectively identifies the patterns based on features selected from improved
GGA. It is accomplished through figuring out the impacting factors that affect fluctuations in solar energy. The
pseudocode for deep CNN is provided as Pseudocode 1.

Pseudocode 1. Deep convolutional neural network

Input: Number on KDD CUP 999 Data Source in IoT
1: Attack Manipulation Function (Kdd Cup 99 Data Source)
2: Confusion Matrix⇐ Kdd Cup 99 Data Source

Global solar energy estimation using improved greedy based genetic algorithm … (Tintu Beena Devaraj)
638  ISSN: 2088-8694

3: Extract Features from Confusion Matrix

4: State Training Data Set (0,1) & Testing Dataset (0,1)
5: Set Sequential Deep Learning Model
6: If Kdd Cup 99 Data Set =1 Then
7: Collect 0s and 1s Information Classification
8: AM ⇐ Sequential Deep Learning Model
9: End If
10: Training Data Set, AM_i ← Data Set
11: If Training Data Set =1, Then
12: Attack Manipulations (AM_i )⇐ Testing Data Set
13: If AM_i = 1, Then
14: Re-Train the Deep Learning Model
15: Else
16: Set Data Verification and Validation Phase
17: End If
18: End If
19: Sources of Classification Data ⇐ AM_i
20: End If End

From Pseudocode 1, the functionality of the algorithm is indicated with steps in which initially it takes
the input data source from the atmosphere containing weather factors that produce solar energy. The variable
AMi is the sequential DL model for training and testing. The training method is operated by adjusting the
internal parameters so that the model is correlated to the output with respect to the input. It reduces the training
error to its maximum when new input data regarding energy flow are fed into the system. An over-fitting
complication is considerably reduced by updating the algorithm by assigning minimized weights. Complex
associations within the global-solar energy features impacting continuous supply are being mapped in order to
achieve optimality in learning new feature data entering the system and aid in the extraction of beneficial data.
Validation and verification of the classification technique yield minimal error and predict important features
disrupting the functionality of systems and classifies the future prediction of energy. Predictions are being
observed from solar energy sources. Solar radiation are certain parameters that fluctuate the supply of energy
to systems, which must be predicted and based on demand and supply is to be maintained for effective energy
systems. The algorithm for deep CNN is projected in Algorithm 2.

Algorithm 2. Deep-convolutional neural network

1. Input Dataset (N samples)
2. ∈← 10−8 , γvi ← 0.9999, γmi ← 0.9
3. αoi = 0.01
4. ρ = θri sw1 {1,2, ∞}m
5. θo ← N(0, √2/Ni
6. t i , mi , vi ← 0
7. t i = t i + 1
8. g it ← ∇θ Li (θti −1 − αti γmi , mti − 1 or Sw2 {0,1}
9. mit ← γmi mt − 1 + (1 − γmi )g it
10. vit ← γvi vt − 1 + (1 − γvi )g it
11. θt = θt1 − ∇ti

From Algorithm 2, it is clear that classification with deep CNN is involved with numerous layers in
the NN and follows a feed-forward based operation. Back-propagation adjusts the learning parameters, such as
biases and weights of the network, in order to minimize the cost value. It performs effective classification by
analyzing complex and larger data by passing it on to multiple neuron layers. Feature maps or filters in the
convolutional layer classify the energy-patterns by adaptively learning spatial features from lower to higher
level features. From the algorithm, the variable ∈ refers to the learning rate and αoi specifies the momentum
parameter. The damping ratio is determined by ρ and the process of updating the performance criteria by
variable θo . Predictions at different time intervals are represented with mit , and γmi refers to the stability
constant. The speed of movement in parameter space is represented by variable vit . Classification based on
significant parameters in solar energy predicts the normal operating characteristics and differentiates the energy
patterns observed from the dataset.

Int J Pow Elec & Dri Syst, Vol. 16, No. 1, March 2025: 633-641
Int J Pow Elec & Dri Syst ISSN: 2088-8694  639

4. RESULTS AND DISCUSSION

The results obtained using the proposed deep CNN model is evaluated to determine the accuracy and
performance efficiency rate of the model. Further, the efficiency of the proposed framework is estimated
using metrics such as MSE, accuracy, MAE and RMSE by doing so the effectiveness of the proposed model
can be identified.

4.1. Performance analysis

The performance of the proposed system is validated with loss metrics such as MSE, MAE, and RMSE
through which benefits of the proposed feature selection and classification algorithm can be determined.
Simulation in MATLAB and outcomes achieved from the present prediction model are depicted in Figure 2. From
Table 1, it is clear that accuracy achieved through the model is found to be 99.75 and sensitivity of 100. These
results indicate the robustness of the proposed model in the prediction of energy consumption. The comparison
of the error values obtained from present algorithm with the traditional algorithm is tabulated in Table 2.
From Table 2, it is clearly indicated that error values calculated from ANN, FL, and GRNN is
comparatively higher than the proposed model. It is due to the modified GGA and deep CNN, the percentage
of error measured in the proposed work is relatively lower (2.023) than the existing methodologies (GRNN –
3.79). The utilization of advanced algorithms such as the improved GGA for feature selection and Deep CNN
for classification indicates a step toward more accurate and efficient energy management systems. The
graphical representation of the comparison is represented in Figure 2.
Figure 2 shows that error calculation performed with existing and proposed indicates that ANN has
the comparatively higher error and the proposed algorithm exhibits its efficiency in terms of MAPE through
lower loss values in the prediction process. In addition, another comparison of existing models in terms of R
and RMSE is being performed and tabulated in Table 2. From Table 3, it is clear that the proposed mechanism
(21% RMSE) achieves lower error values compared to existing empirical (326.8%) and ANN (113.6%)
algorithms. The graphical depiction of the comparison is presented in Figure 3.

Table 1. Performance analysis of deep CNN Table 2. Comparative analysis of proposed work
algorithm using MATLAB with existing algorithm [24]
Performance metrics Obtained values Model Error percentage
Accuracy 99.75 GRNN error 3.79
Sensitivity 100 FL error 3.99
precision 99.2806 ANN error 4.95
Recall 100 Proposed 2.023

Figure 2. Comparative assessment [24]

Figure 3 shows that the proposed algorithm determines the efficiency through reduced error than
empirical and ANN model in predicting the energy consumption pattern. Comparison of measurements in terms
of RMSE and R determines the competence of the proposed algorithm. The comparison of the existing methods
proposed with respect to various error metrics is recorded in given Table 3.
From Table 4, inference from the comparison specifies that proposed research (RMSE 0.215) is
efficient in predicting future energy consumption with reduced loss obtained when compared with different
existing models such as LSTM (0.9), CNN (0.98), and multistep CNN stacked LSTM (0.36). The graphical
representation of the comparative analysis is projected in Figure 4. From Figure 4, a comparative analysis of
various existing methods like LSTM, CNN, and Multistep CNN-based stacked LSTM, it can be noted that the

Global solar energy estimation using improved greedy based genetic algorithm … (Tintu Beena Devaraj)
640  ISSN: 2088-8694

proposed algorithm improved GGA with deep CNN exhibits greater prediction efficiency with minimal loss
than other conventional algorithms. In the error measured by each method, proposed model proved to be least
erroneous. Thus, it confirms the efficacy and reliability of the proposed model.

Table 3. Performance Table 4. Comparing performance efficiency of existing with proposed

comparison [25] methodology [22]
Model R RMSE (%) Methods RMSE MAE R2 MSE MAPE
Empirical model 0.9304 326.8 LSTM 0.9 0.4 0.9 0.82 9.2
ANN model 0.9744 113.6 CNN 0.98 0.78 0.87 0.99 14.48
Proposed 0.986 21.5 Multistep CNN stacked LSTM 0.36 0.18 0.98 0.13 3.11
Proposed 0.215 0.652 0.986 0.0423 2.11

Figure 3. Comparative graphical analysis [25] Figure 4. Comparative performance assessment [22]

5. CONCLUSION
The proposed prediction algorithm employed improved GGA for selecting features and performed the
classification process by deployment of deep-CNN. Experimentation with MATLAB as a primary tool for
analyzing outcomes in predicting the hourly, daily and monthly future solar energy patterns using the proposed
algorithm have shown minimum prediction loss. The performance has been evaluated through metrics such as
MSE, MAE, and RMSE with values as 0.423, 0.652, and 0.215, respectively. The achieved accuracy of 99.75
shows the efficacy and precision of 99.28 suggests the better reliability of the proposed model when compared
to existing models such as CNN (RMSE 0.98, MAE 0.78, MSE 0.99) and LSTM (RMSE 0.9, MAE 0.4, MSE
0.82). Sensitivity and recall at 100 obtained by performance analysis using MATLAB shows improved
efficiency of the proposed classification model. The prediction loss acquired was also compared with other
traditional algorithms which also confirm the robustness of the proposed model, making it a potential tool
system for achieving enhanced prediction accuracy, efficiency, and sustainability in energy management
practices. By integrated techniques and cutting-edge algorithms, the research paves the way for advancements
in the field of energy management and holds promise for addressing critical issues related to distribution and
energy consumption. As a result, the proposed future prediction of solar energy was considered efficient since
it achieved reduced error values than other prediction algorithms. It assists in maintaining stability in solar
energy-based systems. As future work, the present research was focus on predicting, additional weather
parameters which hinders solar energy generation and supply to maintain a stable operational condition and
investigating the scalability of the proposed methodology to real-world energy systems with larger datasets can
provide understandings into its practical applicability. Additionally, conducting comparative studies with other
state-of-the-art algorithms can help validate the effectiveness.

REFERENCES
[1] E. Sarmas, E. Spiliotis, V. Marinakis, G. Tzanes, J. K. Kaldellis, and H. Doukas, “ML-based energy management of water pumping systems
for the application of peak shaving in small-scale islands,” Sustainable Cities and Society, vol. 82, 2022, doi: 10.1016/j.scs.2022.103873.
[2] P. Agee, L. Nikdel, and S. Roberts, “A measured energy use, solar production, and building air leakage dataset for a zero energy
commercial building,” Scientific Data, vol. 8, no. 1, 2021, doi: 10.1038/s41597-021-01082-8.
[3] J. Spriet and A. McNabola, “Decentralized drain water heat recovery: A probabilistic method for prediction of wastewater and
heating system interaction,” Energy and Buildings, vol. 183, pp. 684–696, 2019, doi: 10.1016/j.enbuild.2018.11.036.
[4] S. Qiu, F. Feng, Z. Li, G. Yang, P. Xu, and Z. Li, “Data mining based framework to identify rule based operation strategies for
buildings with power metering system,” Building Simulation, vol. 12, no. 2, pp. 195–205, 2019, doi: 10.1007/s12273-018-0472-6.

Int J Pow Elec & Dri Syst, Vol. 16, No. 1, March 2025: 633-641
Int J Pow Elec & Dri Syst ISSN: 2088-8694  641

[5] Z. Yang, M. Mourshed, K. Liu, X. Xu, and S. Feng, “A novel competitive swarm optimized RBF neural network model for short-
term solar power generation forecasting,” Neurocomputing, vol. 397, pp. 415–421, 2020, doi: 10.1016/j.neucom.2019.09.110.
[6] A. Heidari and D. Khovalyg, “Short-term energy use prediction of solar-assisted water heating system: Application case of combined
attention-based LSTM and time-series decomposition,” Solar Energy, vol. 207, pp. 626–639, 2020, doi: 10.1016/j.solener.2020.07.008.
[7] A. Mellit, A. M. Pavan, and V. Lughi, “Deep learning neural networks for short-term photovoltaic power forecasting,” Renewable
Energy, vol. 172, pp. 276–288, 2021, doi: 10.1016/j.renene.2021.02.166.
[8] M. Rizwan and P. Chaudhary, “Short term solar energy forecasting using GNN integrated wavelet-based approach,” International
Journal of Renewable Energy Technology, vol. 10, no. 3, p. 229, 2019, doi: 10.1504/ijret.2019.10022867.
[9] J. Song, S. D. Oh, and S. J. Song, “Effect of increased building-integrated renewable energy on building energy portfolio and energy
flows in an urban district of Korea,” Energy, vol. 189, 2019, doi: 10.1016/j.energy.2019.116132.
[10] Y. Wei et al., “Prediction of occupancy level and energy consumption in office building using blind system identification and neural
networks,” Applied Energy, vol. 240, pp. 276–294, 2019, doi: 10.1016/j.apenergy.2019.02.056.
[11] N. T. Ngo, “Early predicting cooling loads for energy-efficient design in office buildings by machine learning,” Energy and
Buildings, vol. 182, pp. 264–273, 2019, doi: 10.1016/j.enbuild.2018.10.004.
[12] R. K. Sahu, B. Shaw, J. R. Nayak, and Shashikant, “Short/medium term solar power forecasting of Chhattisgarh state of India using
modified TLBO optimized ELM,” Engineering Science and Technology, an International Journal, vol. 24, no. 5, pp. 1180–1200,
2021, doi: 10.1016/j.jestch.2021.02.016.
[13] M. Mishra, P. Byomakesha Dash, J. Nayak, B. Naik, and S. Kumar Swain, “Deep learning and wavelet transform integrated
approach for short-term solar PV power prediction,” Measurement: Journal of the International Measurement Confederation, vol.
166, 2020, doi: 10.1016/j.measurement.2020.108250.
[14] Mustaqeem, M. Ishaq, and S. Kwon, “A CNN-Assisted deep echo state network using multiple Time-Scale dynamic learning
reservoirs for generating Short-Term solar energy forecasting,” Sustainable Energy Technologies and Assessments, vol. 52, 2022,
doi: 10.1016/j.seta.2022.102275.
[15] H. Bezerra Menezes Leite and H. Zareipour, “Six Days Ahead Forecasting of Energy Production of Small Behind-the-Meter Solar
Sites,” Energies, vol. 16, no. 3, 2023, doi: 10.3390/en16031533.
[16] K. Murugaperumal, S. Srinivasn, and G. R. K. D. Satya Prasad, “Optimum design of hybrid renewable energy system through load
forecasting and different operating strategies for rural electrification,” Sustainable Energy Technologies and Assessments, vol. 37,
2020, doi: 10.1016/j.seta.2019.100613.
[17] K. Konis, “A circadian design assist tool to evaluate daylight access in buildings for human biological lighting needs,” Solar Energy,
vol. 191, pp. 449–458, 2019, doi: 10.1016/j.solener.2019.09.020.
[18] S. Boubaker, M. Benghanem, A. Mellit, A. Lefza, O. Kahouli, and L. Kolsi, “Deep Neural Networks for Predicting Solar Radiation
at Hail Region, Saudi Arabia,” IEEE Access, vol. 9, pp. 36719–36729, 2021, doi: 10.1109/ACCESS.2021.3062205.
[19] J. F. Rushman, P. Thanarak, S. Artkla, and P. Maneechot, “Electrical power demand assessment of a rural community and the
forecast of demand growth for rural electrification in Ghana,” International Energy Journal, vol. 19, no. 3, pp. 177–188, 2019.
[20] I. Jebli, F. Z. Belouadha, M. I. Kabbaj, and A. Tilioua, “Prediction of solar energy guided by pearson correlation using machine
learning,” Energy, vol. 224, 2021, doi: 10.1016/j.energy.2021.120109.
[21] D. Pattanaik, S. Mishra, G. P. Khuntia, R. Dash, and S. C. Swain, “An innovative learning approach for solar power forecasting using
genetic algorithm and artificial neural network,” Open Engineering, vol. 10, no. 1, pp. 630–641, 2020, doi: 10.1515/eng-2020-0073.
[22] N. Elizabeth Michael, M. Mishra, S. Hasan, and A. Al-Durra, “Short-Term Solar Power Predicting Model Based on Multi-Step
CNN Stacked LSTM Technique,” Energies, vol. 15, no. 6, 2022, doi: 10.3390/en15062150.
[23] H. Ali-Ou-Salah, B. Oukarfi, K. Bahani, and M. Moujabbir, “A New Hybrid Model for Hourly Solar Radiation Forecasting Using
Daily Classification Technique and Machine Learning Algorithms,” Mathematical Problems in Engineering, vol. 2021, 2021, doi:
10.1155/2021/6692626.
[24] M. Sridharan, “Generalized Regression Neural Network Model Based Estimation of Global Solar Energy Using Meteorological
Parameters,” Annals of Data Science, vol. 10, no. 4, pp. 1107–1125, 2023, doi: 10.1007/s40745-020-00319-4.
[25] M. Rezakhani, N. Sarrafzadeh-Ghadimi, R. Entezari-Maleki, L. Sousa, and A. Movaghar, “Energy-aware QoS-based dynamic virtual
machine consolidation approach based on RL and ANN,” Cluster Computing, vol. 27, no. 1, 2024, doi: 10.1007/s10586-023-03983-2.

BIOGRAPHIES OF AUTHORS

Tintu Beena Devaraj is an assistant professor from the Department of Electrical and
Electronics Engineering working in Kerala Technological University, Kerala, India. She has 8 years
of teaching and research experience in artificial intelligence. Her main research interests include
performance optimization using machine learning and deep neural networks. She has also a few
publications in the field of total quality management using fuzzy logic. She can be contacted at email:
[email protected].

Charu Navaneeth Raveendran Kottoor is a Mechanical Engineering tutor in the

Applied Technology Department at the Applied Technology School, ACTVET, in Al Ain, United
Arab Emirates. He has 14 years of experience in education at both the college level and technology
high school level. His major areas of interest include Additive Manufacturing and Computer-
Integrated Manufacturing. He has published research papers in the field of robotic cell manufacturing
and additive manufacturing. He can be contacted at email: [email protected].

Global solar energy estimation using improved greedy based genetic algorithm … (Tintu Beena Devaraj)