기계학습을 활용한 공장 지붕형 태양광 발전 구조물의 안전성 평가

Abstract

This paper proposes a machine learning model to solve the time cost of repeated design changes and structural reviews by experts due to the instability of the initial design by non-experts in the construction of rooftop photovoltaic. 4 machine learning models were evaluated for their performance in predicting the safety of photovoltaic. Grid search, Standard scaler, and Over sampling were used to improve the performance of the machine learning models. The model was trained with K-fold cross-validation and evaluated with F1-Score. The RF model showed the best performance with 85% in the safe class and 94% in the unsafe class. The KNN model performed poorly in the safe class due to data imbalance. The SVM model performed well, but required the longest time for prediction at 1.94 seconds. The LR model had the lowest performance. Dimensionality reduction models were also evaluated using PCA, SVD, and ICA. The dimensionality reduction models mostly performed worse than the original models. We analyzed why the models performed worse when dimensionality was reduced. The cumulative variance ratio and information loss error were 99% and 0.021, indicating no data loss. We analyzed the linearity of the data structure and found that most of the Pearson's r coefficients were close to zero, and the correlation between input and output variables was analyzed as nonlinear using surface plots. Due to the non-linear data structure, the dimensionality reduction model performs poorly, with the LR model showing the lowest performance of the four models. Therefore, we propose the RF model as a method for non-structural experts to judge the safety of rooftop photovoltaic structures.