Model Tuning and Performance Evaluation in Machine Learning Models for PV Power Forecasting - Case Study of a BIPV System in Switzerland

Aug 3, 2026

Upload by Asher

High PV penetration can create grid congestion and operational challenges when photovoltaic systems are not integrated effectively. Accurate PV power forecasting supports reliable grid operation and optimized PV management. This work examines how machine-learning hyperparameters shape model structure, learning dynamics, and forecast accuracy for a building-integrated PV system in Switzerland. KNN and SVR are trained using six years of measurements, tuned via grid search, then assessed with K-fold cross-validation to quantify performance differences.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

AccessFree to download

Introduction

Background and Motivation

Machine Learning and Hyperparameter Optimization

Methodology

Model Setup (LR, SVR, KNN)

Performance Evaluation

2024 IEEE 52nd Photovoltaic Specialist Conference (PVSC) ©2024 IEEE DOI: 10.1109/PVSC57443.2024.10748823| 978-1-6654-6426-0/24/$31.00 | Model Tuning and Performance Evaluation in Machine Learning Models for PV Power Forecasting: Case Study of a BIPV System in Switzerland Elham Shirazi a and Hugo Quest b,c a Faculty of Engineering Technology, University of Twente, 7522NB Enschede, The Netherlands, b ´Ecole Polytechnique Fdrale de Lausanne (EPFL), Institute of Electrical and Micro Engineering (IEM), Photovoltaics and Thin-Film Electronics Laboratory, CH-2002 Neuchtel, Switzerland c 3S Swiss Solar Solutions AG, CH-3645 Thun, Switzerland Abstract—High PV penetration into the electricity grid can lead to issues such as congestion if PV systems are not integrated effectively. Accurate PV power forecasting helps to address this issue. In recent years machine learning approaches have gotten a lot of attention in PV power forecasting due to their ability to extract complex relationships between different variables. Hyperparameters are a vital part of machine learning models, influencing their structure, learning process and accuracy of the forecast. Choosing the right hyperparameters is often oneof the key challenges in developing effective machine learning models and therefore accurate PV power forecasts. These hyperparameters can be optimized through various methods including grid search and random search. In this study, two machine learning models including KNN and SVR are used to forecast the power of a BIPV system installed in Switzerland, using 6 years of measurements. Then the grid search has been applied to these models for hyperparameters optimization. Afterwards, the performances of the models are evaluated using K-Fold crossvalidation. The results of this study show that choosing the right hyperparameters leads to a more accurate forecast, as they influence the structure, learning process and performance of the model. I. INTRODUCTION The adoption of Photovoltaic (PV) generation is a key component of the global transition towards a carbon-neutral energy system. Due to many regulations, policies, incentive programs, technology advancement and environmental concerns, the implementation of PV systems has been significantly increased [1] . This can create challenges such as voltage regulation and congestion in the operation of the electricity grid, if PV systems are not integrated effectively [2] . PV power forecasting plays a crucial role in the efficient integration of solar energy into the electricity grid and in optimizing the operation of PV systems. The main challenges in PV forecasting include dealing with the variability in weather conditions, accurately modelling complex nonlinear relationships, and integrating diverse data sources such as satellites, sky imagers or on-site measurements [3] . Various approaches are taken in PV forecasting, including physics-based and machine learning (ML) [4] . Physics-based models are based on physical principles of the atmosphere while machine learning models ignore these physical principles and work with the data. There are of course hybrid models that combine different approaches, often integrating machine learning with physical to leverage the strengths of each. Machine learning has become an increasingly important tool in PV power forecasting, offering significant improvements in accuracy and reliability compared to physics-based forecasting methods. In the context of ML, a hyperparameter is a parameter which is set before the learning process begins. Unlike model parameters, which are learned during training, hyperparameters are not learned from the data but are rather used to control the learning process. Moreover, there is a range of hyperparameters in many ML models where they may interact in nonlinear and complex ways. Hyperparameters play a significant role in the behaviour and performance of a machine learning model, therefore hyperparameter optimizati

FAQs

Why is PV power forecasting important for electricity grid operation?

Accurate forecasting helps integrate solar energy efficiently and mitigates grid issues such as voltage regulation problems and congestion when PV systems are not integrated effectively.

Which machine learning models are used for the BIPV power forecasting case study?

The study uses KNN and SVR for tuning and evaluation, with Linear Regression also mentioned as a baseline model without hyperparameters.

How are hyperparameters optimized and model performance evaluated in the study?

Hyperparameters for SVR and KNN are optimized using grid search, and model performance is evaluated using K-fold cross-validation to compare forecasting accuracy under different model structures.

Aug 3, 2026

Upload by Asher