Explainable AI For Power Plant Boiler Temperature Forecasting And Anomaly Detection

Afandi, Nur (2023) Explainable AI For Power Plant Boiler Temperature Forecasting And Anomaly Detection. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6022212007-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2025. Download (9MB) | Request a copy

Abstract

Explainable AI, also well known as XAI is a method to overcome AI black box problem, which is incapability of an AI model to be explained. This research conducts experiment to implement XAI methods in a forecasting model and anomaly detection model. Heat distribution in a power plant boiler is chosen as case study. In particular, the model architecture implemented is encoder-decoder, the model objectives are forecasting boiler reheater pipes temperature and anomaly detection, and the XAI method is based on feature importance. The research output is information about the most important input variables which is has significant impact to the model output. The forecasting models are based on RNN and Attention. The RNN-based model uses encoder-decoder architecture with stacked LSTM in each layer. The most feasible RNN-based model uses 64 hidden and got RMSE equal to 4.414. Meanwhile, the Attention-based model is based on TFT. The model performances will degrade as the forecasting horizon (decoder length) increases. However, the degraded performance is still acceptable with RMSE below 10 degrees. The anomaly detection is determined based on the 99 th percentile of Mahalanobis Distance distribution as the threshold. The explanation of input variables’ importance is addressed using the ante-hoc and post-hoc methods. The post-hoc method is based on LIME which gives a good visualization but, unfortunately, cannot handle forecasting models with 3-dimensional output. The ante-hoc method is based on an intrinsic feature of the TFT model by utilizing the variable selection network. The explanations give the features importance percentage of input variables, static covariates, and known future inputs.
=====================================================================================================================================
Explainable AI atau lebih banyak dikenal dengan sebutan XAI adalah sebuah metode untuk mengatasi permasalahan kotak hitam pada model AI. Yaitu sebuah permasalahan dimana sebuah model AI tidak transparan dan tidak bisa dijelaskan korelasi antara inpu-output nya. Pada riset ini dilakukan serangkaian percobaan untuk mengimplementasikan metode XAI pada model forecasting dan deteksi anomali. Persebaran panas pada boiler PLTU dijadikan sebagai studi kasus. Secara khusus, model yang diaplikasikan menggunakan arsitektur encoder-decoder, digunakan untuk melakukan forecasting temperatur dari pipa-pipa reheater pada boiler dan untuk mendeteksi terjadinya anomali. Metode XAI yang digunakan berbasis pada feature importance. Output dari penelitian ini adalah informasi mengenai variabel input yang paling signifikan mempengaruhi output dari model AI. Model forecasting
didasarkan pada RNN dan atensi. Model berbasis RNN menggunakan arsitektur encoder-decoder dengan LSTM bertingkat di setiap layer. Model berbasis RNN yang paling sesuai adalah model yang menggunakan 64 hidden unit dan didapatkan nilai RMSE sebesar 4,414. Sementara itu, model berbasis atensi menggunakan arsitektur TFT. Performa model ini akan menurun saat horison forecasting (panjang decoder) meningkat. Namun, penurunan performa masih dapat diterima karena nilai RMSE masih di bawah 10 derajat. Deteksi anomali ditentukan berdasarkan persentil ke-99 dari distribusi jarak
Mahalanobis yang digunakan sebagai nilai ambang batas. Interpretasi tentang feature importance variabel input dibahas dengan menggunakan metode ante-hoc dan post-hoc. Metode post-hoc menggunakan LIME yang memberikan visualisasi yang baik akan tetapi tidak dapat menangani model dengan luaran 3 dimensi. Metode ante-hoc didasarkan pada fitur intrinsik yang ada pada model TFT dengan memanfaatkan jaringan pemilihan variabel. Hasil interpretasi memberikan nilai presentase feature importance variabel input dan kovariat statis.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Forecasting, XAI, Encoder-Decoder, Anomaly Detection
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	Nur Afandi
Date Deposited:	27 Jul 2023 13:24
Last Modified:	27 Jul 2023 13:25
URI:	http://repository.its.ac.id/id/eprint/99641

Actions (login required)

View Item