Machine Learning for performance prediction in smart buildings - photovoltaic self-consumption and life cycle cost optimization

Aug 5, 2026

Upload by Himbo

Photovoltaic systems in buildings are expanding quickly to support clean energy and decarbonization targets, yet increasing PV self-consumption remains challenging because it depends on building–grid interaction. The study addresses time-consuming, multi-objective ESS design by testing 24 machine learning surrogate models for PV self-consumption, including models for short-term thermal energy storage (TES). Results identify top-performing algorithms and show that TES can raise PV self-consumption when electric heat pumps supply heating, cooling, and domestic hot water, while TES sizing optimized by life cycle cost achieves 7.1% savings over 30 years.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages32

Reading Time1 h 21 min

AccessFree to download

Introduction

Background and motivation

Metrics and related indicators

Methodology (surrogate modeling and TES evaluation)

Machine learning models for PV self-consumption

Short-term TES performance assessment

Results and optimization

Model accuracy, training, and prediction speed

TES impact on PV self-consumption

Life cycle cost-based TES size optimization

Highlights and contribution

Machine Learning for performance prediction in smart buildings: photovoltaic self-consumption and life cycle cost optimization Abstract The application of Photovoltaic (PV) system in buildings is growing rapidly in response to the need for clean energy sources and building decarbonization targets. Nonetheless, enhancing PV self-consumption through technical solutions such as Energy Storage Systems (ESS) is getting higher importance to increase the profitability of PV plants, by minimizing the buildinggrid interaction. In this context, analyzing PV self-consumption of different energy storage configurations becomes more relevant and crucial in building energy modeling although it is heavily time-consuming and complicated, particularly within a multi-objective optimization related to the ESS design. As a solution to resolve this issue, this paper evaluates the accuracy, training, and prediction speed of 24 Machine Learning (ML) models to be used as surrogate models for analyzing PV self-consumption in smart buildings. Furthermore, the performance of short-term Thermal Energy Storage (TES) to increase PV self-consumption is assessed and presented using ML models. The results showed the Gaussian Process Regression (GPR), Neural Networks (NN) including bilayered and trilayered NN models, Support Vector Machines (SVM) including the fine gaussian and cubic SVM models, and Ensembles of Trees (EoT) as superior ML models. The results also revealed that TES systems can efficiently increase PV self-consumption in the building equipped with electric heat pumps to provide heating, cooling, and domestic hot water. Moreover, the TES size optimization regarding the Life Cycle Cost (LCC) showed that the LCC-based optimum TES size can yield 7.1% savings within 30 years of the building service life. The novelties of this research are first to provide a reference to select the most suitable ML models in predicting PV self-consumption, second to implement Machine Learning for analyzing the performance of short-term thermal energy storage to enhance PV self-consumption in buildings, and third to carry out an LCC-based optimization on TES size using ML-based prediction models. Keywords: PV self-consumption, Prediction, Machine Learning, Accuracy, Thermal energy storage, LCC Highlights: 1. Machine Learning is known as a reliable technique to predict the energy performance of buildings. 2. The accuracy, training, and prediction speed of 24 ML models are evaluated and compared in this study. 3. Top-ten most accurate ML models in predicting PV self-consumption are introduced by the obtained results. 4. The TES system is an effective solution for improving PV self-consumption in buildings. 5. The LCC-based optimum size of TES leads to 7.1% saving in the total life cycle cost of the building case study. | Nomenclature
AI
COP
DHW
EER
EoT
ESS
GPR
HP
LR
LCC
MAE
ML | Artificial Intelligence Coefficient of Performance Domestic Hot Water Energy Efficiency Ratio Ensembles of Trees Energy Storage System Gaussian Process Regression Heat Pump
Linear Regression
Life Cycle Cost
Mean Absolute Error Machine Learning | MSE
NNR2
RE
RES
RMSE
RT
SVM
TES Ƞdistribution Ƞregulation Ƞemission | Mean Squared Error
Neural Networks
Coefficient of Determination Relative Error
Renewable Energy Sources Root Mean Square Error Regression Trees
Support Vector Machines Thermal Energy Storage Efficiency of distribution systems Efficiency of regulation systems Efficiency of emission systems | | --- | --- | --- | --- | 1. Introduction Within the last few years, the share of Photovoltaic (PV) systems to supply electricity has been rapidly growing provoked by building and industrial decarbonization goals (Shukhobodskiy & Colantuono, 2020) (Gallego-Castillo et al., 2021) . Moreover, the significant decline in the market price of PV systems (López Prol & Steininger, 2020) leads to its large-sca

FAQs

Why is PV self-consumption important for photovoltaic plants in buildings?

Improving PV self-consumption is the main driver of profitability in grid-parity contexts. It increases economic returns and supports environmental performance and grid stability by reducing peak injections and grid-supplied electricity use.

How does the study evaluate machine learning models for PV self-consumption?

It compares 24 machine learning models as surrogate models, focusing on accuracy, training behavior, and prediction speed. The study reports that several models outperform others, including GPR, specific neural network variants, SVM variants, and tree ensembles.

What is the effect of short-term thermal energy storage and its optimal sizing?

Short-term TES systems can efficiently increase PV self-consumption in buildings equipped with electric heat pumps. An LCC-based optimization of TES size yields 7.1% savings over a 30-year service life for the case study.

Aug 5, 2026

Upload by Himbo