Forecasting Solar Irradiance Using Machine

2023 IEEE 64th International Scientific Conference on Information Technology and Management Science of Riga Technical University (ITMS) | 979-8-3503-7029-4/23/$31.00 ©2023 IEEE | DOI: 10.1109/ITMS59786.2023.10317687

Learning Methods
Meerah Karunanithi Ali Al Sajed Braitea Arslan A. Rizvi
University of Europe for Applied University of Europe for Applied School of Int. Manu. & Cont. Engg
Sciences Sciences Qilu Institute of Technology
Potsdam, Germany Potsdam, Germany Jinan City, China
[email protected] [email protected] [email protected]

Talha Ali Khan

University of Europe for Applied
Sciences
Potsdam, Germany
[email protected]*

Abstract— Solar generation must be integrated into the energy sources. The solar energy the planet receives from the
power system to meet rising energy demand and reduce sun each hour is sufficient to supply all of humanity's energy
dependency on fossil fuels. The intermittent nature of solar needs for a year [3]. This urges companies to quickly switch
energy output threatens the grid's stability and efficient
functioning. Accurate solar power forecasting is essential to to renewable energy sources as their primary energy source.
ensure the proper operation of the grid and the suitable In recent years, machine learning has been used in practically
placement of storage resources. In this work, we thoroughly all fields of knowledge. Machine learning is now widely used
investigate machine learning-based short-term solar irradiance in the renewable energy sector to give grid operators and
forecasting. We assess how well several models—including
farmers accurate and timely information regarding solar
gradient boosting regression, random forest, and KNN—
perform using data from Izmir, Turkey. Our findings irradiance projection for the future [4]. Solar energy systems
demonstrate that machine learning approaches outperform connected to grids have caused an unheard-of upheaval in the
conventional time series models, with the KNN and random power industry. Yet, the intervention faces difficulties, such
forest models generally demonstrating the highest performance. as intermittent, impacting energy supply and demand
The suggested model may be applied practically in solar power
plants and expanded to anticipate power production by management. It would be simpler for power grid operators to
incorporating pertinent input variables. Our research sheds take remedial measures in advance to ensure failsafe results
light on the potential of machine learning methods for during periods of rapid or cascading declines in solar energy
forecasting solar irradiance. It emphasizes the significance of if solar irradiance could be forecast with high precision [5].
precise forecasting for the reliability and efficient functioning of
In general, the study seeks to advance knowledge about
the electricity grid. Future studies can examine various feature
engineering techniques and determine how machine learning the potential of ML for solar energy forecasting and offer
models perform over extended periods. insights into the significant variables influencing prediction
accuracy. Solar irradiation forecasts are essential for power
grid management, energy storage systems, balancing energy
Keywords—Solar Irradiance, Machine Learning, Renewable,
Predication, Energy Internet supply and demand, energy efficiency, economic advantages,
and climate change mitigation. By lowering the risk involved
I. INTRODUCTION with investing in solar projects and raising their profitability,
Energy is a commodity that is taken for granted solar energy may provide economic benefits, and keeping
nowadays. Power generators, utilities, and independent track of variations in sun irradiation can aid in mitigating
system operators all agree today that the only way for these climate change's effects on solar energy production. The
scattered resources to collectively offer a stable, dispatchable successful application of solar energy as a critical energy
generating capacity to the energy market is through source depends on accurate sun irradiation forecasts.
sophisticated forecasting, communications, and control [1]. Although there are various models for calculating
Even while using non-renewable energy is at an all-time high different components of solar irradiance[6], these models do
and will soon run out, it harms the environment. not consider any uncertainty inherent to solar irradiance. This
Recently, the development of the energy internet was work investigates the efficacy of machine learning (ML)
proposed, and the global research community is working on techniques for predicting solar energy production, especially
different aspects of developing such a network. Such a using meteorological information like solar irradiance.
network's features include transmission, communication, Accurate solar irradiance forecasting is essential for power
metering and routing, network topology, grid monitoring and system planning and efficient management, especially for
data processing. Moreover, three subsystems have been realizing an efficient energy internet. This becomes
specified for this network [2]: information, network, and especially important considering the growing integration of
energy. Information plays a crucial role in the reliable renewable energy sources like solar power in existing
operation of such an energy network. electrical grids. Machine learning algorithms improve the
Solar energy emerges as a significant clean energy source accuracy of solar irradiance forecasting by incorporating
as the globe searches for new environmentally acceptable weather data, specifically solar irradiance, into the

979-8-3503-7029-4/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Rajshahi University Of Engineering and Technology. Downloaded on May 28,2024 at 11:38:24 UTC from IEEE Xplore. Restrictions apply.
 forecasting models. It will involve a literature review, data irradiance and how to optimize the performance of solar
collection and analysis, and implementation of machine energy systems under different weather conditions. The curve
learning algorithms to compare different models. is characterized by a narrow base, reflecting the shorter
The Solar Radiation and Meteorological Dataset is a period of sunlight during winter, providing solar energy
collection of meteorological and solar radiation parameters systems with less opportunity to harvest energy.
used in research and applications related to solar energy,
atmospheric science, and climate modelling. It includes
temperature, pressure, wind speed, relative humidity, surface
albedo, solar zenith angle, DNI, DHI, clear sky DHI, and
GHI. DNI is a crucial parameter for solar energy systems, as
it represents the amount of solar radiation directly received
by a surface perpendicular to the sun's rays. Forecasting DNI
can be challenging due to various uncertain meteorological
and environmental factors, the most important of which is the
cloud patterns. Accurate forecasts of DNI can have
significant implications for solar energy systems, climate
modelling, and weather forecasting. Fig.2. Graph presenting DNI vs Time (Jan.5 2017).

II. METHODOLOGY III. DATA CORRELATION

Machine learning and statistical modelling techniques can Solar irradiance is a critical factor in the performance and
be used to develop a predictive model that can accurately efficiency of solar energy systems. It is affected by various
forecast DNI at a given location and time. Data pre- meteorological variables, such as temperature, pressure, wind
processing was essential for ensuring the quality and speed, and humidity. This section will explore the correlation
reliability of the results obtained from the analysis. The data between these meteorological variables and solar irradiance
used in the study was generally clean. Still, the time using the Pandas library to gain insights into the factors that
information was stored in separate columns and was influence solar irradiance and develop strategies for
rearranged to the standard DateTime format for easy analysis. optimizing performance.
A high-resolution time series dataset containing
meteorological and solar radiation data was collected from
the beginning of 2017 until the end of 2019 in Izmir, Turkey.
It can train and validate the machine learning models for solar
power forecasting. It can aid in developing more efficient and
cost-effective solar energy systems. The visualization of the
dataset in Fig shows the DNI values vs. time for July 5, 2017.
The line chart shows that the DNI value starts at midnight,
the first data point. It increases as the sun rises, peaks around
midday, and decreases until sunset. The smooth line graph
indicates no significant fluctuations or disturbances in solar
irradiance throughout the day. The time stamps every 15
minutes provide a detailed record of solar irradiance over the
entire day and can be used to identify patterns and trends to Fig.3. Presenting the correlation between DNI and meteorological data.
optimize solar energy systems. Such a type of data represents
a clear day with no clouds. A correlation heat map is a visual depiction that shows the
degree and direction of the correlation between variables in a
data collection. It can assist researchers in selecting the most
crucial variables, eliminating unnecessary variables, and
locating potential confounding variables. It is an effective
visualization technique to enhance comprehension and
interpretation of complicated data sets [5].

Fig.1. Graph presenting DNI vs. Time (July 5, 2017).

Another visualization of DNI vs time for January 5, 2017,

is shown in Fig. This graph provides valuable insights into
the impact of weather, i.e., cloud coverage, on solar

Authorized licensed use limited to: Rajshahi University Of Engineering and Technology. Downloaded on May 28,2024 at 11:38:24 UTC from IEEE Xplore. Restrictions apply.
 less of the Earth's atmosphere. As a result, there is a greater
DNI since less direct sunlight is dispersed or absorbed, and
more of it reaches the Earth's surface.
The correlation values between features and the DNI, the
goal variable, can be used to select predictors for DNI
forecasting in machine learning models. However, accuracy
is not guaranteed, and model performance should be assessed
using cross-validation. Feature engineering techniques like
feature transformation or feature creation can be applied to
improve model performance.

IV. MACHINE LEARNING MODELS AND IMPLEMENTATION

DETAILS
Random Forest (RF), Gradient Boosting (GB), Support
Vector Machines (SVM), Artificial Neural Networks (ANN),
K-Nearest Neighbors (KNN), and Long Short-Term Memory
(LSTM) are all popular machine learning algorithms that can
be used for various applications. RF and GB are collective
Fig.4. Heat map presenting correlation of DNI with meteorological
features. learning methods that combine multiple decision trees to
improve predictive accuracy and reduce overfitting. SVM is
The correlation values indicate a relationship between a robust algorithm for classification and regression tasks that
each feature and the target variable, DNI. The values of each seeks to find the best hyperplane that separates the data points
feature's linear association with the target variable, DNI, into different classes or predicts the output variable [7]. ANN
show its intensity and direction. Temperature and DNI have is a popular deep-learning algorithm inspired by the human
a somewhat positive association, with a value of 0.514912. brain's structure and function. KNN is a simple and effective
Relative humidity and solar zenith angle have a substantial algorithm that finds the K nearest neighbours of a new data
negative connection with DNI. In contrast, surface albedo, point and assigns it the label or value of most of its
pressure, and wind speed have a moderate negative neighbours [7]. LSTM is a powerful deep learning algorithm
association. These characteristics are not significant for sequential data analysis that can model long-term
predictors of DNI [6]. dependencies and memory.
The ratio of water vapour in the air to the maximum
A. Results and Discussion
amount it can contain at a given temperature is known as
relative humidity. The air can contain more water vapour than The coefficient of determination (R2) was used to assess
at that temperature when the relative humidity rises. This rise the effectiveness of various models, comparing predicted
in water vapour can potentially lower the quantity of DNI solar irradiance values to actual values. Results showed that
KNN with RF kernel had the highest R2 value compared to
reaching the surface by scattering and absorbing sunlight—
all other techniques. The coefficient of determination, or R-
the direct quantity of solar radiation (DNI) from the sun that squared value, is determined by dividing the sum of squares
reaches Earth's surface. When the air is dry, DNI is less from the regression by the sum of all squares. In contrast to
impacted by light scattering and absorption. A drop in DNI the regression sum of squares (SSR), which is the sum of the
occurs when the air is humid because the water vapour in the squared differences between each predicted value and the
atmosphere may scatter and absorb more sunlight. mean of all values, the total sum of squares (SSTO) is the sum
As a result, as relative humidity rises, more water vapour of the squared differences between each observed value and
is present in the air, which enhances sunlight scattering and the mean of all values. The difference in squares between
absorption while decreasing DNI. On the other hand, as each observed value and its matching anticipated value is
relative humidity falls, less water vapour is present in the air known as the erroneous sum of squares (SSE).
to scatter and absorb sunlight, increasing DNI. R2 = SSR / SSTO (1)
The angle between the sun and the zenith, or the point 2
In addition, R can also be expressed as one minus the
directly overhead, is known as the solar zenith angle. As the quotient of the error sum of squares and the total sum of
solar zenith angle rises, the sun travels toward the horizon and
squares.
away from the zenith. The more extended trip that the light (2)
R2 = 1 - SSE / SSTO
from the sun must take through the Earth's atmosphere when
the solar zenith angle rises increases the likelihood that the The degree to which a model predicts the event correctly
light will be scattered and absorbed. DNI drops because less is shown by the coefficient of determination or R2. The
direct normal irradiance (DNI) reaches the Earth's regression line entirely fits all data points when the R2 value
surface. On the other hand, as the solar zenith angle drops, is 1, and the model's predictions are identical to the observed
the sun is nearer, and the light must traverse the zenith and variance. The model does not adequately explain the variance
of the data when the R2 value is near 0, which results in

Authorized licensed use limited to: Rajshahi University Of Engineering and Technology. Downloaded on May 28,2024 at 11:38:24 UTC from IEEE Xplore. Restrictions apply.
 inaccurate predictions. It suggests that the model does not suit Comput. Conf. IWCMC 2019, pp. 1458–1462, 2019, doi:
10.1109/IWCMC.2019.8766480.
the data and may need more optimization or alternative [5] C. N. Obiora, A. Ali, and A. N. Hassan, "Predicting Hourly Solar
modelling strategies. The model performs worse the closer Irradiance Using Machine Learning Methods," 11th Int. Renew. Energy
the R2 value approaches 0. Congr. IREC 2020, no. Irec, pp. 3–8, 2020, doi:
10.1109/IREC48820.2020.9310444.
[6] S. A. Al-Sanea, M. F. Zedan, and S. A. Al-Ajlan, "Adjustment factors
TABLE I. R2 COEFFICIENT VALUE COMPUTED FOR EACH MACHINE for the ASHRAE clear-sky model based on solar-radiation
LEARNING MODEL measurements in Riyadh," Appl. Energy, vol. 79, no. 2, pp. 215–237,
Random Gradient 2004, doi: 10.1016/j.apenergy.2003.11.005.
Metric SVM ANN KNN LSTM
Forest Boosting [7] F. Jawaid and K. NazirJunejo, "Predicting daily mean solar power using
machine learning regression techniques," in 2016 Sixth International
R2 Score 0.8945 0.8400 0.8824 0.8273 0.9100 0.7881
Conference on Innovative Computing Technology (INTECH), 2016, pp.
355–360, doi: 10.1109/INTECH.2016.7845051.

The KNN model, with an R2 score of 0.91, has performed

the best among all the other models, indicating that it
accurately predicts the solar irradiance values based on the
input features. The Random Forest model has also performed
well, and the SVM model has performed well. The Gradient
Boosting and ANN models perform moderately, while the
LSTM model performs poorly. Overall, the KNN and
Random Forest models are the most promising for predicting
solar irradiance based on the given features.

V. CONCLUSIONS
The successful implementation of a solar energy
forecasting model can significantly impact the renewable
energy sector by improving energy management, reducing
costs, and minimizing environmental impact. This study
Implemented machine learning techniques for solar
irradiance forecasting in Izmir, Turkey. Results showed that
machine learning models, especially KNN and Random
Forest, performed better on average. Still, feature engineering
and numerical weather prediction can improve accuracy.
Future research should focus on comparing different machine
learning algorithms and exploring the use of ensemble
models to improve accuracy.
This study highlights the benefits of accurate solar energy
forecasting for grid stability and cost-effective operation.
Advanced machine learning techniques like KNN and
Random Forest outperform traditional time series models.
Future research can focus on developing new and more
accurate models, integrating multiple data sources, long-term
forecasting, and investigating the impact of climate change
on forecasting. These advancements can potentially improve
the accuracy and reliability of energy forecasting, benefiting
the renewable energy industry and broader society.

REFERENCES
[1] J. Li, J. K. Ward, J. Tong, L. Collins, and G. Platt, "Machine learning
for solar irradiance forecasting of photovoltaic system," Renew. Energy,
vol. 90, pp. 542–553, 2016, doi: https://doi.org/10.1016/j.renene
(Djaafari A., 2022)2015.12.069.
[2] H. M. Hussain, A. Narayanan, P. H. J. Nardelli, and Y. Yang, "What is
energy internet? Concepts, technologies, and future directions," IEEE
Access, vol. 8, no. iv, pp. 183127–183145, 2020, doi:
10.1109/ACCESS.2020.3029251.
[3] A. Moosa, H. Shabir, H. Ali, R. Darwade, and B. Gite, "Predicting Solar
Radiation Using Machine Learning Techniques," Proc. 2nd Int. Conf.
Intell. Comput. Control Syst. ICICCS 2018, no. Iciccs, pp. 1693–1699,
2019, doi: 10.1109/ICCONS.2018.8663110.
[4] A. Javed, B. K. Kasi, and F. A. Khan, "Predicting solar irradiance using
machine learning techniques," 2019 15th Int. Wirel. Commun. Mob.

Authorized licensed use limited to: Rajshahi University Of Engineering and Technology. Downloaded on May 28,2024 at 11:38:24 UTC from IEEE Xplore. Restrictions apply.