Electronic Theses and Dissertation

Penelitian ini menerapkan algoritma bayessian regularization neural network (BRNN) dalam membuat model prediksi beban listrik. Penerapan algoritma BRNN dimaksudkan untuk mencegah kondisi overfitting yang kerap terjadi pada penerapan algoritma neural network dalam membuat model prediksi. Kondisi overfitting mempengaruhi tingkat akurasi model prediksi sehingga nilai prediksi yang diinginkan tidak mendekati nilai sebenarnya. Dalam penelitian ini, model prediksi BRNN dibangun dengan 5 (lima) topologi jaringan dengan skema supervised learning feed forward network. Kelima topologi jaringan model BRNN tersebut dilatih dan diuji dengan menggunakan konstruksi dataset yang berbeda. Selain itu, model BRNN dibandingkan dengan model prediksi levenberg marquardt neural network (LMNN) dengan skema topologi jaringan yang sama. Hasil penelitian menunjukkan, model BRNN yang dibangun dengan 5 topologi jaringan tersebut mampu menghasilkan model prediksi dengan tingkat akurasi yang tinggi berdasarkan simulasi pelatihan model dan pengujian model prediksi

This study applies the Bayessian regularization neural network (BRNN) algorithm in creating an electrical load prediction model. The application of the BRNN algorithm is intended to prevent overfitting conditions that often occur in the application of neural network algorithms in creating prediction models. Overfitting conditions affect the level of accuracy of the prediction model so that the desired prediction value is not close to the actual value. In this study, the BRNN prediction model was built with 5 (five) network topologies with a supervised learning feed forward network scheme. The five network topologies of the BRNN model were trained and tested using different dataset constructions. In addition, the BRNN model was compared with the Levenberg Marquardt neural network (LMNN) prediction model with the same network topology scheme. The results of the study showed that the BRNN model built with the 5 network topologies was able to produce a prediction model with a high level of accuracy based on simulations of model training and testing of the prediction model.