Bipolar lead-acid batteries offer advantages over traditional valve-regulated lead-acid batteries including compact design, shorter current path, better active mass utilization, and higher power density. Since cathode and anode are on opposite sides of the bipolar substrate, electrons directly traverse to adjacent cells without requiring external circuit elements such as tabs, straps, and posts. A study published in Frontiers of Chemical Science and Engineering presents a novel approach for estimating the state of health of these advanced batteries.
The research team fabricated six 6V bipolar lead-acid battery prototypes using fused filament fabrication for acrylonitrile butadiene styrene components. Spot-welded multilayered lead foils served as bipolar substrates for pasting positive and negative active materials. Batteries underwent cycle tests at 25 degrees Celsius, charged at 0.3 A and discharged at 0.3 A until 5.25 V cutoff, repeated until state of health fell below 60 percent.
The study utilized partial charging profiles to extract three health feature attributes: localized voltage area calculated from the voltage-time curve between 6.45 V and 6.70 V, sample entropy quantifying uncertainty of sequential voltage data, and fuzzy entropy measuring complexity. Gray relational analysis validated that all attributes strongly correlate with battery health, with gray relational grades exceeding 0.83.
The hybrid estimation framework integrates Lasso regression and support vector regression as first-stage models, whose outputs combine as input for a random forest regression model optimized by gray wolf algorithm. The gray wolf optimizer fine-tuned hyperparameters including number of trees, maximum depth, and minimum samples to split a node.
The hybrid models utilized two health attribute pairs: localized voltage area with fuzzy entropy and localized voltage area with sample entropy. Training data came from batteries BLAB01 to BLAB04, testing from BLAB05 and BLAB06. For localized voltage area with fuzzy entropy, the hybrid model achieved average mean absolute error below 1.02 percent and average root mean squared error below 1.5 percent. Relative error values remained under 6.2 percent despite irregular deviations, with 88 percent below 3.5 percent.
The hybrid model was compared with Gaussian process regression, deep neural network, recurrent neural network, and long short-term memory models, demonstrating slightly better overall performance. Long-term estimation using 50 percent to 70 percent initial training data showed root mean squared error decreasing from 2.08 percent to 1.27 percent as training data increased.
This research validates that the gray wolf optimized hybrid regression framework provides accurate state of health estimation for bipolar lead-acid batteries using only partial charging profiles, enabling timely replacement or preventive maintenance.
DOI
10.1007/s11705-025-2613-7