STATE OF CHARGE ESTIMATION DEVELOPMENT OF BATTERY ENERGY STORAGE SYSTEM WITH DOMAIN INFORMED NEURAL NETWORK

• Main Author: Henry Widatyoko, Andrean
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73101
• Institution: Institut Teknologi Bandung

<p align="justify">The use of renewable energy in microgrid systems can be increased by scheduling the charging and discharging of the battery energy storage system. To be able to schedule optimally, an accurate State of Charge (SOC) estimation by the battery management system is required. In general, these estimations are categorized into two methods: modeling based on the principles of physics and data-driven modeling. Physics rule-based modeling requires extensive domain knowledge, rigorous feature engineering, and long modeling time; this makes the method unsuitable for estimating SOC in real time. Meanwhile, data-driven modeling is relatively faster by utilizing measured signals (such as voltage, current, and temperature) and is able to find complex data patterns, even though this modeling category is agnostic to applicable physics rules. Therefore, the modeling in this study uses the Domain Informed Neural Network (DINN) method as a combination of physics-based and data-driven modeling so that the two can complement each other.This research is the initial stage in the development of a Digital Twin (DT) battery energy storage system from the microgrid of the Energy Management Laboratory. Furthermore, the SOC modeling that has been done can be developed and used in the preparation of battery energy storage system charging and discharging scheduling algorithms. Therefore, this research focuses on developing the SOC model of Valve Regulated Lead Acid (VRLA) batteries installed in battery energy storage system using the DINN method.The DINN method is equipped with domain knowledge in the form of constraints, namely approximation constraints and monotonicity constraints, which support the accuracy and compliance of the model with the applicable physics principles in SOC modeling. This is an advantage as well as a differentiator between DINN and other methods. Based on the model evaluation, the performance of the DINN method was obtained with the following performance metrics: Root Mean Square Error (RMSE) 0,301%; Mean Absolute Error (MAE) 0,231%; Mean Absolute Percentage Error (MAPE) 0,407%; SOC biggest prediction of 99,5%; and SOC smallest prediction of 3,1%. Unlike the case with the performance of the SOC model with the Deep Neural Network (DNN) method (as a comparison), which has performance metrics of RMSE 0,41%; MAE 0,32%; MAPE 0,699%; SOC biggest prediction of 102,5%; SOC smallest prediction of 2,6%. This evaluation shows that the prediction performance of the SOC model with the DINN method is better than the performance of the DNN model with predicted SOC values in the range of 0-100%.