Prediction of Selected Parameters of Energy Storage System using Recurrent Neural Networks

Research on Ultracapacitors in Hybrid Systems: Case Study

“…Examples of such difficult conditions are low ambient temperature, long time gaps between engine operation phases, significant power draw from vehicle onboard electrical appliances (i.e., the dashboard), and in the case of vehicles featuring start-stop systems [13], high frequency of system operation (frequent combustion engine start-ups and short periods of battery charging-dynamic conditions that cause the battery to lose charge at a faster rate, decreasing the usable capacity and durability [40]). Additional drawbacks of batteries are self-discharge and difficulty in detecting the discharge point of the battery (often, full discharge occurs suddenly, with no warning [41]). According to the scientific literature [42][43][44][45][46][47][48][49][50], the rate of self-discharge can range from 5% to even 30% SOC over one month, depending on the type of battery and storage conditions (i.e., ambient temperature) [51][52][53].…”

“…It is worth remembering that the SOC estimation error is significant because dynamic processes are taken into consideration, such as voltage relaxation. More accurate SOC estimation (even during load) can be achieved through tools, such as various combinations of the Kalman Filter [21,[59][60][61] and other prediction methods [69] based on recurrent neural networks [41,70].…”

Comparison of NARX and Dual Polarization Models for Estimation of the VRLA Battery Charging/Discharging Dynamics in Pulse Cycle

“…Currently, on the Polish market, the price for 1 kWh for electrochemical batteries is within 60-150 euro/kWh [23][24][25][26] (e.g., lead-acid batteries or Valve-Regulated Lead-Acid (VRLA) Absorbent Glass Mat (AGM) batteries) and more than 300 euro/kWh (Li-ion battery) [26]. Due to this, VRLA AGM batteries have found a wide range of applications in the industry [27][28][29][30][31][32][33][34][35][36], e.g., in forklifts where VRLA batteries are used as ballast, in start-up of gas engines [27], Start and Stop systems in vehicles with conventional drivetrains [15] and cogeneration systems where the gas engines 2 of 28 have been employed [9], as well as other distributed energy generation systems and emergency power supply systems connected with auxiliary electricity generators (e.g., in hospitals).…”

“…In the second part of the work, a Nonlinear AutoRegressive eXogenous model (NARX) in the form of the recurrent artificial neural network (R-ANN) was used to the estimation of the voltage on the VRLA battery terminals. Mostly, ANN models simplify the identification procedure with high accuracy [30,69]. They have distinct advantages over linear identification methods, i.e., the approximation of multivariable nonlinear functions, the simple gradient-based adaptation of model parameters and a rapid calculation of ANN equations.…”

“…In particular, interesting for our research are Recurrent Artificial Neural Networks (R-ANN) where connections between units form a directed cycle. Previous research [30] shows, that with properly chosen and pre-processed data, system parameters, and operating conditions, recurrent ANN model can be used for prediction, to some extent, for energy storage performance under a wide variety of operating conditions. Considered in this work, NARX model based on R-ANN plays a significant role because it can be used as a predictor, nonlinear filter and as a nonlinear model of the dynamical system which can be later used in model-based control algorithms [113,114] in the BMS systems.…”