Study of Charge–Discharge Energy Prediction Using Neural Networks for Improving the Operational Economic Efficiency of Lithium-Ion Battery Aggregation

Abstract

In recent years, the supply of renewable energy has increased, and the destabilization of electric power systems is being caused by the fluctuation and time-maldistribution of the renewable energy output. To solve this problem, battery aggregation was studied. Lithium-ion battery module is a strong candidate for aggregated rechargeable batteries. It is known that the degradation phenomenon, which leads to the increase in the loss of charge–discharge energy (CDE) will occur in a lithium-ion battery. This will cause an economic loss of the battery aggregation. Each of the aggregated lithium-ion battery modules shows degradation states individually and diversely according to the type and operative conditions (e.g., Temperature and charge–discharge power) of lithium-ion battery modules. Using the battery modules preferentially, which have good CDE loss characteristics, leads to the improvement of the operational economic efficiency of battery aggregation. Therefore, it is important to recognize the characteristics of CDE loss of the lithium-ion battery modules individually by the degradation diagnosis. In the past, we derived the equation for CDE, which contains full-charge capacity (FCC), mean-working voltage (MWV), and internal impedance (Z), as shown in equation (1). MWV is the mean voltage of the charge and discharge, and it is similar to the open-circuit voltage. The image of FCC, MWV, Z, and CDE is shown in Fig.1, and these three variables are defined in the equations (2)—(4). In addition, we estimated the coefficient vector Kn of the equations (2)—(4) using equation (5). VVDFD in equation (5) is the voltage difference, as shown in Fig. 2. The symbol tVDFD in Fig.2 is the discharge span to extract the VVDFD value from the charge–discharge cycle data. We investigated this method by applying the data of charge–discharge cycles, which was carried out with the conditions listed in Table.1. It showed good accuracy, and CDE was predicted within 1% error from the estimated Kn and equations (1)—(5) [1]. However, the constants in equation (5) should be calculated using many charge–discharge cycle data as the teacher data to use this method. In this case, more than 350 cycles data were required for good CDE prediction accuracy. Moreover, if there are batteries that have unknown and complex relationship between VVDFD and Kn , the estimation of Kn will be difficult. We tried to predict CDE by using neural networks instead of the Kn estimating process. We used neural networks with three layers as shown in Fig.3. The input values were three batches of data, including discharge voltage, discharge current, state of charge, and bias parameter. This batch data are the values that can be easily acquired from the battery management system (BMS) such as VVDFD . The number of neurons of the hidden layer was adjusted according to the prediction accuracy. The output values were three batches data. The output batch data of MWV and Z were defined as the values corresponding to the regular intervals state of charge (k) value of 50 pieces. We studied some activation functions (e.g., sigmoid, hyperbolic, and power). Estimation accuracy of each parameter was dependent on the number of hidden layer neurons and the activating function. The best conditions of the number of hidden layer neurons and activation function are summarized in Table.2. The estimation results of MWV and Z were approximated as the sixth polynomial of k, and these approximated polynomials and estimated FCC were substituted in equation (1) to calculate CDE. Teacher data could be set to less than 350 cycles. When tVDFD was set to 10 seconds, and teacher data were set to 100 cycles data, CDE prediction was carried out with the error ratio of 0.25% (charge) and 0.14% (discharge) on average, as shown in Fig.4. The accuracy improved compared with the method of coefficient vector estimation by voltage difference, which was carried out with the error ratio of 0.28% (charge) and 0.41% (discharge) on average. It is suggested that CDE can be predicted from the unknown and complex relationship between indices from BMS data and Kn by using neural networks. [1] M. Arima, L. Lin, and M. Fukui, “Three Degradation Parameters Estimation of a LIB Module Using Single Indicator for In-situ Charge–Discharge Energy Prediction”, in Proc. IEEE International Telecommunications Energy Conference, Torino, Italy, Oct., 2018. Figure 1