Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Short-Term Photovoltaic Power Forecasting Using Machine Learning in Data-Scarce Environments: A Case Study in Bhutan

Carreira, Ricardo LU (2025) MVKM05 20251
Department of Energy Sciences
Abstract
This thesis proposes a multi-step methodology for short-term photovoltaic (PV) power
forecasting in regions lacking extensive historical PV generation data, with a case study
conducted in Dewathang, Bhutan.
As Bhutan’s energy security faces increasing threats from climate change—primarily due
to its heavy reliance on hydropower—there is growing national interest in diversifying energy
sources, particularly through the deployment of solar PV. However, effective integration
of large-scale PV systems requires accurate forecasting techniques, typically dependent on
long-term historical data of both meteorological conditions and PV output.
This study situates its contribution within Bhutan’s broader sustainability and renewable
energy... (More)
This thesis proposes a multi-step methodology for short-term photovoltaic (PV) power
forecasting in regions lacking extensive historical PV generation data, with a case study
conducted in Dewathang, Bhutan.
As Bhutan’s energy security faces increasing threats from climate change—primarily due
to its heavy reliance on hydropower—there is growing national interest in diversifying energy
sources, particularly through the deployment of solar PV. However, effective integration
of large-scale PV systems requires accurate forecasting techniques, typically dependent on
long-term historical data of both meteorological conditions and PV output.
This study situates its contribution within Bhutan’s broader sustainability and renewable
energy goals. It provides an overview of PV technology fundamentals and reviews state-ofthe-
art machine learning (ML) methodologies for solar irradiance and PV power forecasting.
The proposed approach combined weather data and newly collected PV output data, applying
ML-based forecasting and regression models to enable 1-hour ahead PV power predictions.
The best-performing models for 1-hour ahead meteorological forecasting achieved R2
scores of 0.9073 for solar irradiance, 0.9815 for temperature, and 0.9460 for relative humidity.
In the regression task for PV power based on collected, the top model reached an R2 of
0.989.
Although the final integration of the forecasting and regression components was constrained
by the lack of more recent data, this work demonstrates the feasibility and effectiveness
of ML-based forecasting in data-scarce contexts. The proposed methodology has
significant implications for enabling reliable PV power predictions in regions with low levels
of PV deployment, but historical weather data. (Less)
Popular Abstract
As different nations aim to build energy systems that are simultaneously climate-resilient, technologically innovative, and environmentally friendly, Bhutan stands as an example for the world, with a national energy mix that is almost entirely composed of renewable energy. Nevertheless, its dependence on hydropower generation has prompted Bhutanese institutions to explore alternative energy sources, such as solar energy.
Solar energy can be harnessed using technologies such as photovoltaic (PV) panels. However, to assess the viability of PV production projects at different scales, PV power generation must be accurately predicted. Machine learning (ML) algorithms can help perform this task—but how can such tools, which rely heavily on... (More)
As different nations aim to build energy systems that are simultaneously climate-resilient, technologically innovative, and environmentally friendly, Bhutan stands as an example for the world, with a national energy mix that is almost entirely composed of renewable energy. Nevertheless, its dependence on hydropower generation has prompted Bhutanese institutions to explore alternative energy sources, such as solar energy.
Solar energy can be harnessed using technologies such as photovoltaic (PV) panels. However, to assess the viability of PV production projects at different scales, PV power generation must be accurately predicted. Machine learning (ML) algorithms can help perform this task—but how can such tools, which rely heavily on long-term historical data, be used in regions where no previous PV installations or measurements exist?
This degree project proposes a methodology to predict PV power production in regions with limited PV generation data, provided that historical meteorological data are available. The methodology was tested through a case study conducted in Dewathang, southern Bhutan.
The methodology follows several steps. First, it reviews PV technology basics and modern machine learning approaches for solar forecasting. Then, it combines long-term historical weather records with newly collected local PV generation data. In the first stage, machine learning models predict key meteorological variables—such as solar irradiance, temperature, and humidity—one hour ahead. These forecasts achieve high accuracy (R² up to 0.98). In the second stage, regression models use weather data to estimate PV power output, reaching an R² of 0.989.
Although limited recent data prevented full system integration, the study demonstrates that reliable one-hour-ahead PV forecasting is feasible even in data-scarce regions, supporting sustainable energy transitions. (Less)
Please use this url to cite or link to this publication:
author
Carreira, Ricardo LU
supervisor
organization
course
MVKM05 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Machine Learning, Photovoltaic, Bhutan, Short-term Forecasting, PV Power
report number
ISRN: LUTMDN/TMPH-25/5624-SE
ISSN
0282-1990
language
English
id
9222776
date added to LUP
2026-02-16 12:47:11
date last changed
2026-02-16 12:47:11
@misc{9222776,
  abstract     = {{This thesis proposes a multi-step methodology for short-term photovoltaic (PV) power
forecasting in regions lacking extensive historical PV generation data, with a case study
conducted in Dewathang, Bhutan.
As Bhutan’s energy security faces increasing threats from climate change—primarily due
to its heavy reliance on hydropower—there is growing national interest in diversifying energy
sources, particularly through the deployment of solar PV. However, effective integration
of large-scale PV systems requires accurate forecasting techniques, typically dependent on
long-term historical data of both meteorological conditions and PV output.
This study situates its contribution within Bhutan’s broader sustainability and renewable
energy goals. It provides an overview of PV technology fundamentals and reviews state-ofthe-
art machine learning (ML) methodologies for solar irradiance and PV power forecasting.
The proposed approach combined weather data and newly collected PV output data, applying
ML-based forecasting and regression models to enable 1-hour ahead PV power predictions.
The best-performing models for 1-hour ahead meteorological forecasting achieved R2
scores of 0.9073 for solar irradiance, 0.9815 for temperature, and 0.9460 for relative humidity.
In the regression task for PV power based on collected, the top model reached an R2 of
0.989.
Although the final integration of the forecasting and regression components was constrained
by the lack of more recent data, this work demonstrates the feasibility and effectiveness
of ML-based forecasting in data-scarce contexts. The proposed methodology has
significant implications for enabling reliable PV power predictions in regions with low levels
of PV deployment, but historical weather data.}},
  author       = {{Carreira, Ricardo}},
  issn         = {{0282-1990}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Short-Term Photovoltaic Power Forecasting Using Machine Learning in Data-Scarce Environments: A Case Study in Bhutan}},
  year         = {{2025}},
}