TEMPERATURE ANOMALY DETECTION IN BATTERY ENERGY STORAGE SYSTEM (BESS) USING MACHINE LEARNING METHODS

• Main Author: Rizki Febrina, Vany
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/56828
• Institution: Institut Teknologi Bandung

Battery Energy Storage System (BESS) provides a way to enhance the durability of the present grid structure. Therefore, a Battery system for safety evaluation is crucial to reduce the safety risk of the battery system to an acceptable level. The BESS in operation can suffer from excessive heat due to internal processes inside the battery. It can lead to hazardous consequences, such as the risk of explosion. Hence, it is necessary to detect the temperature change anomalies within the system. It prevents internal battery faults that can cause performance degradation and reducing battery life. In this work, instead of using the battery temperature, the feature is the rate of battery temperature change. Therefore, the anomaly can be detected earlier. Data-driven methods, especially machine learning methods, are currently widely used for system anomaly detection. The model obtains using the supervised learning method. Supervised learning provides a more accurate and fast learning model by incorporating prior domain knowledge of the system compared to the unsupervised one. The machine learning methods used include Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Regression (RVP). Based on the quantitative assessment (metric model), RF is slightly more accurate than ANN and RVP. However, if the model is used to predict battery temperature, ANN is the best performer with an RMSE value of 0.13 and MAE 0.10. Thus, RMSE and MAE cannot be the primary reference to justify the model performance. Therefore, ANN will be used to estimate the temperature of batteries. ANN provides a learning model that is more accurate than the other two methods. This research conducted using the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology. The Business Understanding phase defines the data mining goal to detect the battery temperature anomaly within the BESS. The BESS consists of 18 modules where each module builds on 15 cells of high capacity of lithium ferro phosphate (LFP) prismatic batteries with nominal voltage 3.2VDC and 100Ah. The Data Understanding phase describes the data quality and the data structure (physical variables and data hierarchy). Before the data enter the modelling phase, the raw data get through the cleaning and filtering process in the Data Preparation phase. The temperature rates of changes feature also generated in this phase as the target feature. The data then used to train the battery temperature model using the supervised learning method in the Modelling phase. The model input features are voltage, current, power, cell position, and temperature operation, while the target feature is the temperature rates of changes. The best model selected among the three algorithms includes Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Regression (SVR). ANN is a best performer to estimate the temperature. In addition, a statistical approach utilized to determine the limit of outlier data on temperature estimation using ANN, an anomaly occurs when the change in battery temperature exceeds the value of three standard deviations with a data confidence interval of 99.7%. The evaluation results will display in the deployment phase. Based on statistical analysis, the test data on 11 November 2019 at 00:00 to 23:59 WIB total 386355 data found anomalies of 0.34% or 1334 data with details of cell 1 of 67 data; 5%, cell 2 of 86 data; 6 %, cell 3 for 98 data; 7%, cell 4 for 82 data; 6%, cell 5 for 80; 6%, cell 6 for 111 data; 8%. cell 7 for 103 data; 8%, cell 8 for 108 data; 8%, cell 9 for 75 data; 6%, cell 10 for 95 data; 7%, cell 11 for 64 data; 5%, cell 12 for 93 data; 7%, cell 13 for 73 data; 6%, cell 14 for 72 data; 5%, cell 15 for 128 data; 10%. Cell 15 has the greatest number of anomalies. Based on cabinet-level calculations, Cabinet 3 found anomalies of 233 out of the total anomalies. Keywords: anomaly detection, battery energy storage system, machine learning, battery temperature, rate of changes