Multi-Site PV Forecasting Across Climate Zones and Prediction Horizons
(2026) In CODEN:LUTEDX/TEIE EIEM01 20261Division for Industrial Electrical Engineering and Automation
- Abstract
- In light of the ongoing energy transition, solar photovoltaic (PV) systems are becoming increasingly widespread. The intermittent nature of solar power introduces challenges to grid operation, which is why PV power forecasting has emerged as an important research topic. While existing research mostly focuses on single-site forecasting, this thesis compares sites from four different Köppen-Geiger climate zones to assess geographical generalisability, alongside prediction horizons relevant to the day-ahead and intraday market.
Six models are compared: Naïve Persistence, Random Forest, XGBoost, Ridge Regressor, SVR and LSTM, with all hyperparameters optimised using Optuna. While differences between models are modest, forecasting accuracy... (More) - In light of the ongoing energy transition, solar photovoltaic (PV) systems are becoming increasingly widespread. The intermittent nature of solar power introduces challenges to grid operation, which is why PV power forecasting has emerged as an important research topic. While existing research mostly focuses on single-site forecasting, this thesis compares sites from four different Köppen-Geiger climate zones to assess geographical generalisability, alongside prediction horizons relevant to the day-ahead and intraday market.
Six models are compared: Naïve Persistence, Random Forest, XGBoost, Ridge Regressor, SVR and LSTM, with all hyperparameters optimised using Optuna. While differences between models are modest, forecasting accuracy varies substantially across climatic regions. The sunnier zones, BWh (Subtropical desert) and Csa (Mediterranean), consistently outperform the more variable BSk (Mid-latitude steppe) and Dfb (Humid continental). Within each zone, accuracy deteriorates during rainier seasons due to increased cloud cover and weather variability.
Regarding forecast horizons, all models outperform Persistence at 15 minutes but underperform it at 24 hours. A Rolling Forecast framework reveals that LSTM forecasts converge towards a generic diurnal PV profile when future weather data is unavailable. A Perfect Rolling Forecast framework using a Bi-LSTM with perfect weather inputs isolates meteorological uncertainty as a primary driver of horizon-dependent accuracy loss.
The findings suggest that climatic context is a more fundamental determinant of forecasting accuracy than model choice, and that meteorological variability drives forecast error regardless of horizon or model complexity - contributing to a more generalised understanding of PV forecasting across diverse climatic conditions. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9241694
- author
- Mc Mullin, Niamh LU
- supervisor
-
- Huan Li LU
- organization
- course
- EIEM01 20261
- year
- 2026
- type
- H3 - Professional qualifications (4 Years - )
- subject
- keywords
- PV, photovoltaic, solar, forecasting, Köppen-Geiger, climate zone, machine learning, forecast horizon, prediction horizon, day-ahead, intraday, optuna
- publication/series
- CODEN:LUTEDX/TEIE
- report number
- 5564
- language
- English
- id
- 9241694
- date added to LUP
- 2026-09-10 09:29:03
- date last changed
- 2026-09-10 09:29:03
@misc{9241694,
abstract = {{In light of the ongoing energy transition, solar photovoltaic (PV) systems are becoming increasingly widespread. The intermittent nature of solar power introduces challenges to grid operation, which is why PV power forecasting has emerged as an important research topic. While existing research mostly focuses on single-site forecasting, this thesis compares sites from four different Köppen-Geiger climate zones to assess geographical generalisability, alongside prediction horizons relevant to the day-ahead and intraday market.
Six models are compared: Naïve Persistence, Random Forest, XGBoost, Ridge Regressor, SVR and LSTM, with all hyperparameters optimised using Optuna. While differences between models are modest, forecasting accuracy varies substantially across climatic regions. The sunnier zones, BWh (Subtropical desert) and Csa (Mediterranean), consistently outperform the more variable BSk (Mid-latitude steppe) and Dfb (Humid continental). Within each zone, accuracy deteriorates during rainier seasons due to increased cloud cover and weather variability.
Regarding forecast horizons, all models outperform Persistence at 15 minutes but underperform it at 24 hours. A Rolling Forecast framework reveals that LSTM forecasts converge towards a generic diurnal PV profile when future weather data is unavailable. A Perfect Rolling Forecast framework using a Bi-LSTM with perfect weather inputs isolates meteorological uncertainty as a primary driver of horizon-dependent accuracy loss.
The findings suggest that climatic context is a more fundamental determinant of forecasting accuracy than model choice, and that meteorological variability drives forecast error regardless of horizon or model complexity - contributing to a more generalised understanding of PV forecasting across diverse climatic conditions.}},
author = {{Mc Mullin, Niamh}},
language = {{eng}},
note = {{Student Paper}},
series = {{CODEN:LUTEDX/TEIE}},
title = {{Multi-Site PV Forecasting Across Climate Zones and Prediction Horizons}},
year = {{2026}},
}