Utilização de redes neurais informadas pela Física no preenchimento de falhas em séries de temperatura do ar

Abstract

In this study, a model based on Physics-Informed Neural Networks (PINN) was developed and implemented with the aim of filling gaps in micrometeorological data and compared with seven other models. The model’s cost function was derived from an empirical version of the molecular diffusion equation. Results were compared with seven traditional machine learning models, including Linear Regression, LASSO Regression, Elastic Net (EN), K-Nearest Neighbors (KNN), Classification and Regression Trees (CART), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP). The results indicated that neural networkbased models, including PINN, outperformed most other models, with an average Root Mean Square Error (RMSE) of 0.08 °C compared to the average RMSE values of 0.12 °C for the other models. PINN also required fewer training epochs compared to MLP, with a maximum difference of only 62 epochs in five of the eight analyzed regions. However, LASSO and EN models exhibited the highest RMSE values, averaging approximately 0.37 °C across all phases and regions. Additionally, the models demonstrated resilience to varying proportions of training data, showcasing their ability to adapt to different data ratios, with all models meeting the minimum requirement for effective learning from the utilized data. In summary, this study highlighted the effectiveness of neural network-based models, including PINN, in predicting temperatures in micrometeorological data, providing a precise solution for gap filling. These results contribute to the advancement of Environmental Sciences and offer insights for future studies in the field of missing data imputation, enhancing the integrity of climatic and environmental analyses.