Comparative Study of Machine Learning on Synthetic

and Experimental Solar Data

Dr. Narjisse Kabbaj and Dr. Tayeb Brahimi

1 Introduction
This project compares the performance of machine learning algorithms when applied to two
different approaches:

1. Using mathematical modeling to generate synthetic data from experimental solar data,
followed by applying machine learning on the synthetic data.

2. Applying machine learning directly on the experimental solar data.

The goal is to determine which method provides more accurate predictions and better
generalization to unseen data. The project will utilize various mathematical models and
machine learning algorithms to evaluate their performance.

2 Workflow
2.1 Experimental Solar Data
The project begins with analyzing real solar data. This data includes multiple features and
a target output variable. The features (X) and target variable (Y) are defined as follows:

• Features (X):

– Date: The time at which the data was recorded.

– Air Temperature (C): The ambient air temperature in degrees Celsius, which
influences the solar energy received and the efficiency of energy conversion.
– Relative Humidity (%): The percentage of moisture present in the air, which
can affect the clarity of the atmosphere and the amount of solar radiation reaching
the surface.
– Barometric Pressure (mB (hPa equiv)): The atmospheric pressure mea-
sured in millibars, equivalent to hectopascals (hPa). This parameter can influence
weather conditions, which in turn affects solar energy reception.

• Target Variable (Y):

1
 – GHI (Wh/m²): Global Horizontal Irradiance (GHI) measured in watt-hours per
square meter. This parameter represents the total amount of shortwave radiation
received from the sun per unit area on a horizontal surface, which serves as our
target output variable for prediction.

These features (X) will be used to predict the target variable (Y), which is the GHI
(Global Horizontal Irradiance). The comparison between applying machine learning on syn-
thetic and experimental solar data aims to provide insights into which approach yields better
predictive accuracy.
The data will be provided in an Excel file and organized for analysis. Students will prepare
the data by handling missing values, normalizing features, and splitting it into training and
testing sets.

2.2 Track 1: Mathematical Modeling and Machine Learning on

Synthetic Data
2.2.1 Mathematical Modeling
In this step, students will develop several mathematical models using the experimental solar
data, such as:

• Linear Model:
y = mx + b
where y is the output, x is the input variable, and a and b are constants.

• Logarithmic Model:
y = a · ln(x) + b
where y is the output, x is the input variable, and a and b are constants.

• Exponential Model:
y = a · ebx
where a and b are constants, and e is the base of the natural logarithm.

• Polynomial Model:

y = an · xn + an−1 · xn−1 + · · · + a1 · x + a0

where n is the degree of the polynomial, and an are the polynomial coefficients.

2.2.2 Error Estimators

The accuracy of each model will be evaluated using the following error estimators:

2
 • Mean Squared Error (MSE):
n
1X
M SE = (yi − ŷi )2
n i=1

where yi is the actual value, ŷi is the predicted value, and n is the number of data
points.

• Root Mean Squared Error (RMSE):

√
RM SE = M SE

• R-squared: Pn
2 (yi − ŷi )2
R = 1 − Pi=1
n 2
i=1 (yi − ȳ)
where ȳ is the mean of the observed values.

The model with the lowest error, based on these estimators, will be selected as the best
model for generating synthetic data.

2.2.3 Synthetic Data Generation

Once the best model is selected, it will be used to generate synthetic data simulating various
conditions, such as different irradiance and temperature scenarios. The goal is to create a
dataset large enough to train and test machine learning algorithms effectively.

2.2.4 Machine Learning on Synthetic Data

After generating synthetic data, students will apply machine learning algorithms, such as:

• Random Forest

• k-Nearest Neighbors (k-NN)

• Artificial Neural Networks (ANN)

These algorithms will be trained on the synthetic dataset and their performance will be
evaluated using metrics such as accuracy and error rates.

2.3 Track 2: Machine Learning Directly on Experimental Data

In this track, students will directly apply the same machine learning algorithms on the
original experimental solar data. The goal is to compare how well machine learning performs
on real-world data without the intermediate step of mathematical modeling. The same
algorithms (Random Forest, k-NN, ANN) will be used, and their results will be compared
to those obtained from the synthetic data.

3
3 Comparison of Results
In the final step, the performance of the machine learning models from both tracks will be
compared. Students will analyze:

• The accuracy of predictions from both synthetic and experimental data.

• The generalization ability of models on unseen data.

• The strengths and limitations of each approach.

The comparison will provide insights into which method—using mathematical modeling and
synthetic data or direct application of machine learning on experimental data—yields better
results.

4 Final Deliverables
4.1 IEEE-Style Article
Students will submit a detailed article in the IEEE format. The article should include:

• Abstract: A concise summary of the project.

• Introduction: The problem statement and project objectives.

4
 • Methodology: A detailed explanation of the mathematical modeling process, ma-
chine learning algorithms, and data analysis.

• Results: A presentation of the results from both synthetic and experimental data
tracks.

• Conclusion: A discussion of the comparison between the two approaches, including

strengths and limitations.

4.2 Presentation
Students will also present their findings in a well-structured presentation, which should
include:

• Introduction: Overview of the project and its objectives.

• Data Description: Explanation of the experimental solar data.

• Mathematical Modeling: A brief overview of the models developed and their per-
formance.

• Machine Learning: Explanation of the machine learning algorithms used and their
results on both synthetic and experimental data.

• Comparison: A comparison of the results from both tracks.

• Conclusion: Final thoughts on the performance of each approach and possible future
improvements.