E3S Web of Conferences 309, 01163 (2021) https://doi.org/10.

1051/e3sconf/202130901163
ICMED 2021

Analysis Of Solar Power Generation Forecasting Using Machine

Learning Techniques
K. Anuradha1,* Deekshitha Erlapally2, G. Karuna3, V. Srilakshmi4, K. Adilakshmi5
1Professor, Computer Science and Engineering, GRIET, Hyderabad, Telangana, India.
2MTech Student, Computer Science and Engineering, GRIET, Hyderabad, Telangana, India.
3Professor, Computer Science and Engineering, GRIET, Hyderabad, Telangana, India.
4 Asst.Professor, Computer Science and Engineering, GRIET, Hyderabad, Telangana, India.
5 Asst.Professor, Computer Science and Engineering, GRIET, Hyderabad, Telangana, India.

Abstract: Solar power is generated using photovoltaic (PV) systems all over the world. Because the output
power of PV systems is alternating and highly dependent on environmental circumstances, solar power
sources are unpredictable in nature. Irradiance, humidity, PV surface temperature, and wind speed are only
a few of these variables. Because of the unpredictability in photovoltaic generating, it's crucial to plan ahead
for solar power generation as in solar power forecasting is required for electric grid. Solar power generation
is weather-dependent and unpredictable, this forecast is complex and difficult. The impacts of various
environmental conditions on the output of a PV system are discussed. Machine Learning (ML) algorithms
have shown great results in time series forecasting and so can be used to anticipate power with weather
conditions as model inputs. The use of multiple machine learning, Deep learning and artificial neural
network techniques to perform solar power forecasting. Here in this regression models from machine
learning techniques like support vector machine regressor, random forest regressor and linear regression
model from which random forest regressor beaten the other two regression models with vast accuracy.

1 Introduction: large quantities of uncontrollable generation at this time.

As it is converting solar radiance       
Solar energy has many benefits, but also have their get that how much power is emitted for different
initial investment for installing solar panels is quite high, location, time, and weather. For this type of clarification
and not everyone will be able to afford them. machine learning techniques are used in order to
Unfortunately, this is a downside of solar panels; differentiate it for different conditions. Machine-learning
nevertheless, as prices continue to decline, the future techniques are wide applied to several fields where it can
looks bright. Solar panels are currently relatively costly; separate the weather based power.
but, new government programs and cutting-edge The amount of energy a PV system generates is
technology are making them cheaper. Despite the fact proportional to meteorological parameters including
that photovoltaic cells are recognized as the significant cloud cover, sun intensity, and site-specific conditions,
source of potential energy production, their low return among other [3]. Solar panel works differently for
on investment and high upfront costs keeps them from different weather conditions. In case if its summer
becoming widely used.The high initial cost prevents seasons then the amount of energy consumed by the
them from becoming widely used. Because photovoltaic panel from sun is very much more. But in case of rainy
cells convert solar energy into electrical energy, the and windy conditions the energy consumed is pretty
amount of solar energy produced each day influences the much different. Power generation mostly depends on
size of the photovoltaic system, just as the amount of weather conditions so they take weather forecasting into
solar radiation influences the amount of electricity consideration. As a result, the amount of electricity
produced each day. This is influenced by factors such as generated is determined by solar irradiance on a given
location, time, and weather patterns. Solar irradiance is day, which is determined by a number of factors such as
the power obtained per unit area from the Sun via location, time, and weather patterns. We concentrate on
electromagnetic radiation in the wavelength range of the the problem of automatically generating models that
solar cell in use. accurately predict renewable generation based on
Major grid integration is difficult because renewable National Weather Service forecasts (NWS). Using
energy is irregular and uncontrollable. Households can historical NWS forecast data and data generated by solar
now use almost any amount of energy due to the recent panels, we experiment with a variety of machine
electric grid at any moment, but it is not equipped for learning techniques to develop prediction models.

*
Corresponding author: [email protected]

© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0
(http://creativecommons.org/licenses/by/4.0/).
E3S Web of Conferences 309, 01163 (2021) https://doi.org/10.1051/e3sconf/202130901163
ICMED 2021

Meteorological data, including ambient temperature, 2. Related Work:

humidity, and solar radiation, will be collected by
meteorological monitoring stations every three hours. Solar energy forecasts can be categorised in a variety of
Machine-learning techniques have been widely used ways. The persistence or smart persistence model, which
in a range of fields involving data-driven problems in uses historical data to forecast future power generation
recent decades. Machine-learning approaches encompass over a short period of time, is the most basic method (2-3
a wide range of interdisciplinary topics, including hours). This method can be used to set a standard against
statistics, mathematics, artificial neural networks, data which other forecasting methods can be measured. In
mining, optimization, and artificial optimization. With or most cases, a prediction is completed in two stages. A
without mathematical problem forms, machine learning NWP is designed for a specified time period and location
approaches attempt to find a relationship between input to begin with. The generated NWP is then utilised to
and output data. The process of analysing data is known forecast power generation using forecasting algorithms.
as data analyzation. ML employs statistical approaches It is possible to employ a physical model, a statistical
          method, or a machine learning methodology [1].For
to be explicitly programmed. Machine learning has two prediction, ML algorithms are compared to the Smart
main application categories: regression and Persistence (SP) approach, with ML models
classification. Solar power forecasting necessitates the outperforming the SP model. The unpredictability of
use of regression methods. Some of the ML regression solar resources has hampered grid management as solar
algorithms that can be used for time series forecasting diffusion rates have increased. Unpredictability and
are Linear Regression (LR), Support Vector Machine intermittent electricity delivery are two of the most
Regression (SVMR), and Random Forest (RF). difficult aspects of integrating renewables into the
Weather and physical elements influence the system. As a result, solar power forecasting is becoming
electrical power output of a solar photovoltaic (PV) increasingly important for grid stability, optimal unit
panel. Solar irradiance, cloud cover, humidity, and commitment, and cost-effective dispatch. To overcome
ambient temperature are the main meteorological factors the problem, we employ machine learning techniques to
that influence solar power generation. Predicted weather sift through extraordinary solar radiation predicting
parameters can be used as model inputs, while solar models. For developing prediction models, a variety of
power forecasts can be used as the model output. regression algorithms are tested, including linear least
Because of its ongoing training nature, the ML algorithm squares and support vector machines with various kernel
adjusts to physical parameters. functions. We use day-ahead sun radiation data forecasts
In machine learning SVM plays a major role in order in these tests to show that a machine learning approach
to classify the data and monitor weather condition can correctly anticipate short-term solar power [2]. A
according. Combining data from photovoltaic power hybrid or mixed forecasting method was developed by
generation with meteorological conditions, according to combining clustering, classification, and regression
the positive position of photovoltaic power generation. approaches to produce a forecasting model. Based on the
For every 3 hours svm gives analyzed data for weather forecast for the next day, the model (with the
classification and regression analysis. Using hyperplane closest weather condition) is chosen to forecast the
we can classify the accurate results from solar panel power output using cluster-wise regression [3].
based on the weather conditions.Random forest, on the Renewable energy sources are progressively being
other hand, is a classification strategy that uses many integrated into electric networks alongside nonrenewable
decision trees to classify data. In order to generate an energy sources, posing significant issues due totheir
uncorrelated forest of trees whose committee forecast is sporadic and erratic nature In order to address these
more trustworthy than that of any single tree, bagging issues, soft-computing solutions for energy prediction
and feature randomization are utilised in the are essential.We apply a number of data mining
development of each individual tree. It gives multiple methodologies, including preparing historical load data
decisions tress it merges all the decision tress into one and analysing the features of the load time series,
     for different climatic conditions because electricity consumption is entangled with the
such as for summer, rainy, winter seasons. Error usage of other energy sources like natural gas and oil.
statistics such as mean bias error (MBE), mean absolute The trends in power consumption from renewable and
error (MAE), root mean square error (RMSE), relative nonrenewable energy sources were examined and
MBE (rMBE), mean percentage error (MPE), and contrasted. A novel machine learning-based hybrid
relative RMSE are used to assess the model's validity technique (SVR) uses multilayer perceptron (MLP) and
(rRMSE). Linear regression is a supervised learning- support vector regression [5].Using SVM regression,
based machine learning approach. It does a regression solar power generation produces acceptable results [6].
analysis. Based on independent variables, regression However, it lacks a detailed examination of solar power
models a goal prediction value. It's generally used in generation and meteorological data, and hence is
forecasting to figure out how variables are related. restricted in its capacity to accurately predict other data
Regression models differ in terms of the sort of link sets by merely using different SVM kernels after some
examined between dependent and independent variables, basic statistical data processing [8].
as well as the number of independent variables used. To study the association between expected weather
conditions and power output created as a historical time
series, artificial intelligence (AI) approaches are applied.

2
E3S Web of Conferences 309, 01163 (2021) https://doi.org/10.1051/e3sconf/202130901163
ICMED 2021

AI approaches use algorithms that can implicitly mean solar irradiance and meteorological data in order to
characterise the nonlinear and highly intricate accurately estimate power generated.
relationship between input data (NWP predictions) and The proposed work's System Architecture is to first
output power instead of formal statistical analysis. The consider the dataset and preprocess the data, then divide
ANN is a brain model that is based on biology. They're it into train and test data, apply classification techniques,
employed in a range of applications that use AI and predict the results. Solar power weather dataset is
approaches including supervised, unsupervised, and used for forecasting purposes in this case. Data
reinforcement learning. The ANN learns from data in the preprocessing methods include cleaning, integration,
supervised learning approach by being trained to reduction, and transmission. We must purge any data
approximate and estimate the function or relationship. that is no longer absolutely necessary. Data cleaning is
[6]. the process of identifying and removing inaccurate or
Their models have been improved to predict PV plant incorrect records from a dataset. Data from the real
power generation [47]. Even with the cloud graph from world frequently contains noise and missing values, and
synchronous meteorological satellites, the significant it may be in an unusable format that cannot be directly
unpredictability in critical components, particularly the used for DL models. Data preprocessing is required to
diffuse component from the sky hemisphere, makes solar clean data and prepare it for various Deep Learning
irradiance far less predictable than temperature. PV models, increasing accuracy and efficiency. Training and
systems including a large number of different tiles testing data are separated from the preprocessed data.
deployed over a large area have additional challenges The model is trained using training data, and its
[12]. Because it is impossible to examine all connected predictions are validated using testing data. Data
meteorological forecasts in a practical context, many splitting is the process of dividing available data into two
alternative alternatives have been devised. Weather halves, usually for cross-validator purposes. The first set
forecasts from meteorological websites [8] were of data is used to build a predictive model, while the
considered by some. Others used nonlinear modelling second set is used to evaluate the model's performance.
approaches like artificial neural networks to try to In analyzing data mining algorithms, separating data into
simplify the solar forecast model (ANN). Two types of training and testing sets is crucial. The training
networks are commonly used to forecast global solar percentage is set at 80% and the test percentage is set at
radiation, solar radiation on titled surfaces, daily solar 20%. When a data set is divided into a training set and a
radiation, and short-term solar radiation: radial basis testing set, the majority of the data is used for training
function (RBF) and multilayer perception (MLP). and only a small portion is used for testing. To train any
In a three-layer feed forward model, back- model, no matter what type of dataset is used, the dataset
propagation is the neural network training technique. To must be divided into training and testing data. The
reduce forecast error, the input layer provides an error dataset will be examined for null values and outliers
correction factor depending on the projected output for during the data preprocessing step, and the model will be
the previous 5 minutes. trained using three hours of data before being used to
An LSTM network will learn a function that accepts forecast solar power generation value. The power
a sequence of previous solar irradiance values as input generated radiance phase will be estimated using
and returns a solar irradiance value as output. Deep machine learning (ML) methods (e.g., support vector
neural networks, such as the Deep Belief Network regression, linear regression, elastic net regression, and
(DBN), will learn a function that takes a sequence of random forest) as shown in below Fig 1.
historical sun irradiance values as input and outputs a
solar irradiance value. If a series of observations are
converted into a variety of occurrences, an LSTM
network can learn from them. The sequence is
partitioned using LSTM for prediction purposes.

3 PROPOSED WORK:
For knowing how much power is generated from solar
we have the dataset showing daily average temperature
in Celsius, distance from solar noon, wind speed, wind
direction, sky cover, and humidity and then the power
generated. Here we are calculating how much power is
generated in different weather condition for India dataset
. We have taken Indian dataset with different
temperature readings. The available dataset is based on
hourly weather parameter values. To convert the data to
mean values per day, the average of the 24-hour data
was used. From 2019 to 2020, several weather factors
were collected to investigate the relationship between

3
E3S Web of Conferences 309, 01163 (2021) https://doi.org/10.1051/e3sconf/202130901163
ICMED 2021

(LR) [10].It uses linear predictor functions to represent

the relationship between the input and output variables,
and a least squares approach is used to estimate the
unknown model parameters from the data. A set of linear
equations or an iterative method like gradient descent
can be used to estimate parameter values. We employed
the characteristics provided in the, followed by feature
scaling, to standardise the input data. SVR's precision
varies depending on the kernel function and other
variables. To discover the optimal settings, we employed
the Grid search approach. To evaluate the models'
performance on the test set, we computed the Root-
Mean-Square Error (RMSE) and R squared values.
Before choosing the models with the lowest root mean
squared error and highest R squared values, we fine-
tuned the model hyperparameters. Nonlinear
relationships can be mapped using these methods. In
data science challenges of various kinds, methods
including decision trees, RF, and gradient boosting are
commonly utilised.The RF method is a tree-based
machine learning approach that can be used for
regression and classification. It also performs
Fig.1 Flow chart for solar power generation dimensional reduction, controls missing and outlier
values, and performs a variety of additional data
exploration activities. The bagging approach is used to
4. Methodology:
train RFs. This method allows for the usage of numerous
The current dataset is based on hourly weather parameter instances for the training stage because the dataset is
values. To convert the data to mean values per day, the sampled with a replacement. Linear regression is a
average of the 3-hour data was used. Various weather method for demonstrating the link between a dependent
characteristics were gathered in order to investigate the variable and one or more independent variables by using
relationship between mean solar irradiance and the best-fit linear curve. It is concerned with determining
meteorological data in order to accurately estimate mean the best-fit line with the data by attaining a perfect slope
solar irradiance. The average daily values of air and intercept value.The best model for forecasting solar
temperature, humidity, wind speed, wind direction, power system output based on numerous weather
visibility, average pressure, average wind speed, and parameters was then created. The models that gave the
electricity generated are among the data collected. The greatest results on the dataset were support vector
direction of the wind, on the other hand, indicates how regression, random forests, and linear regression, and
high the sun is. It's also expressed in degrees. these models were then utilized to anticipate PV system
Machine Learning (ML) Models are used for performance for 2019. Thanks to the predictive analysis,
forecasting the solar power generation weather analysis. the estimated production in this situation ranges from 0
The Regression techniques here proposed are Support to 1000 Watt hours. These models were then evaluated
Vector Machine, Random Forest, Linear Regression are using the test data. The SVR model has an RMSE of
various ML Models used in this paper. 135.7, while the random forest model has an RMSE of
28.62 and the SVR model has an RMSE of 58.24. The
random forest model's points are close to the regressed
4.1 Forecasting models: diagonal line, however the SVR model's points are not.
In this study, we used the chosen dataset to evaluate
individual performance using a number of 5 RESULTS:
meteorological attributes utilising three commonly used
machine learning algorithms. The output of the unseen A matrix of pair correlation coefficients is generated for
test sample is predicted to be the mean of these K closest a set of features under investigation in order to find
matches because our prediction variable is continuous collinear factors as shown in below Fig 1, Fig 2,Fig 3
valued. We investigated a variety of K values, however and Fig 4.
only the results for K=3 and K=5 are presented. When K
is more than 3, the RMS error increases.
Support vector regression (SVR) using a radial basis
function as the kernel and random forest (RF)
approaches are used to create the models. Because of the
non-linearity of the dataset, we used the models
indicated above instead of linear models. The most basic
and widely used regression method is linear regression

4
E3S Web of Conferences 309, 01163 (2021) https://doi.org/10.1051/e3sconf/202130901163
ICMED 2021

Fig 5: winter season analysis

Fig 2 correlation matrix

Fig 6: moderate analysis

Fig 3:summer season analysis

Fig7       

Here fig6 shows that how power is generated

comparatively in different weather conditions where the
x axis shows the power generated and y axis shows the
distance from solar noon. It is nothing more than Solar
noon occurs when the Sun passes through a location's
meridian (a meridian is an imaginary line that runs from
the North Pole to the South Pole along the Earth's
surface.) and ascends to its highest point in the sky. In
most cases, it does not occur at 12 p.m and as shown in
above Fig 5 and Fig 6 and Below Fig 7.

Fig 4: rainy season analysis

5
E3S Web of Conferences 309, 01163 (2021) https://doi.org/10.1051/e3sconf/202130901163
ICMED 2021

 $        temperature,56 to 66C is 2

            !!  &+#  /   
summer season and as shown in below Table 1.
Table 1: Model Prediction of RMSE, MAE, MAPE, and MBE
and Accuracy

Model RMSE MAE MSE Accuracy

Support 131.44 77.16 172.76 -44.92

vector
machine
regressor
Fig 8: Different temperature analysis results of how much Linear 58.57 48.39 343.08 72.4
power is generated. regression
Random 27.32 12.45 746.48 94.01
Fig6 says the power generated is read as Jules here forest
where as temperature as in summer, rainy, winter and the regressor
moderate is said as the different weather condition.
Dataset considered is the numerical dataset so we
replaced them with string type where in dataset it is 6. Conclusion:
shown as 0 for rainy 1 for winter, 2 for moderate,3 for We presented a machine learning-based approach for
summer as shown in below Fig 9 and Fig 10. solar power generation analysis in this paper, which
accurately forecasts power generated across India's states
based on environmental data. Most importantly, our
methodology went beyond prediction by delivering key
results that aided in the understanding of solar power
analysis (variable importance by time period). By a wide
margin, the proposed method outperformed other
popular methods, such as Random forest. The proposed
models are SVR, LR, and RF. Compared to the
temperature with the given data. 56-55F --30% power
generation is increasing compared to other temperatures.
Temperature <46F is 17% temperature average of it. As
the above results, we can see that Random Forest
Regressor model is performing better with 94.01%
accuracy and hence that model is preferred for
deployment.
Fig 9: score of all the three model used in solar power analysis.
Reference:
This fig8 shows that score of three different models
used in this paper for solar power generation 1. P. A. G. M. Amarasinghe and S. K.
Abeygunawardane, "Application of Machine
Learning Algorithms for Solar Power Forecasting
in Sri Lanka" (2nd International Conference On
Electrical Engineering (EECon), Colombo, Sri
Lanka, 87 2018).
2. M. Z. Hassan, M. E. K. Ali, A. B. M. S. Ali and J.
Kumar, "Forecasting Day-Ahead Solar Radiation
Using Machine Learning Approach" (4th Asia-
Pacific World Congress on Computer Science and
Engineering (APWC on CSE), Mana Island, Fiji,
252 2017).
3. A. Bajpai and M. Duchon, "A Hybrid Approach of
Solar Power Forecasting Using Machine Learning"
(3rd International Conference on Smart Grid and
Fig 10: temperature difference and their power generated
percentage accordingly.
Smart Cities (ICSGSC), 108 2019).
4. A. Khan, R. Bhatnagar, V. Masrani and V. B. Lobo,
This shows that how much percentage of power is "A Comparative Study on Solar Power Forecasting
generated ass the temperature rises less than 46 Celsius using Ensemble Learning," (4th International
        !"!# 

6
E3S Web of Conferences 309, 01163 (2021) https://doi.org/10.1051/e3sconf/202130901163
ICMED 2021

Conference on Trends in Electronics and

Informatics (ICOEI), 224 2020).
5. Khan, P.W.; Byun, Y.-C.; Lee, S.-J.; Kang, D.-H.;
Kang, J.-Y.; Park, H.-S. Energies, 13, 4870 (2020).
6. Faquir, Sanaa & Yahyaouy, Ali & Tairi, H. &
Sabor, Jalal. International Journal of Fuzzy System
Applications. 4, 10 (2015).
7. Aler R., Martín R., Valls J.M., Galván I.M.
Intelligent Distributed Computing VIII. Studies in
Computational Intelligence, vol 570 (2015).
8. Y. Wang, G. Cao, S. Mao and R. M. Nelms,
"Analysis of solar generation and weather data in
smart grid with simultaneous inference of nonlinear
time series," (IEEE Conference on Computer
Communications Workshops (INFOCOM
WKSHPS), 600 2015).
9. Carrera B, Kim K. Sensors (Basel). 20, 3129 (2020).
10. Jawaid F, NazirJunejo K. Predicting daily mean
solar power using machine learning regression
techniques. (Sixth International Conference on
Innovative Computing Technology (INTECH) 355
2016).
11. Batcha RR, Geetha MK. A survey on IOT based on
renewable energy for efficient energy conservation
using machine learning approaches. (3rd
International Conference on Emerging Technologies
in Computer Engineering: Machine Learning and
Internet of Things (ICETCE) 123 2020).
12. Li, Zhaoxuan & Rahman, Sm Mahbobur & Vega,
Rolando & Dong, Bing. Energies. 9, 55 (2016).
13. Lai JP, Chang YM, Chen CH, Pai PF. Applied
Sciences; 10, 5975 (2020).
14. Brahma, B.; Wadhvani, R. Symmetry, 12, 1830
(2020).