Statistical and Machine-Learning-Based Durability-Testing Strategies for Energy Storage

As the global climate is heating up at accelerated rates, human civilization is turning away from fossil fuels to renewable energy sources. While abundant, wind, solar, and water power are intermittent, motivating investment in long-duration energy storage technology research. The grand vision is to predict circa 20 years of energy-storage cycling behavior with less than one year of new testing data. On the other hand, machine learning and artificial intelligence are revolutionizing most aspects of science and engineering, but adopting those advancements has been slow in the energy-storage community. The reasons are the expense of battery testing and the associated sparsity of datasets, the need to extrapolate instead of interpolating, the inherent stochasticity of the problem, and the deep connection between failure mechanisms and physical and chemical processes. All of this causes a purely data-driven approach to be suboptimal. Herein we propose a new customization of a Gaussian Process for stochastic, domain-knowledge-informed energy storage-failure-distribution prediction. For that, we equip the Gaussian Process with tailored prior mean and kernel functions to give it the ability to use experts’ domain knowledge of candidate failure mechanisms to extrapolate into the future while minimizing the amount of required data. Our results show that our method can, in fact, approximate failure distributions early in the testing process. In short, this work provides the basis for revolutionizing energy storage testing and discovering new technologies for energy storage systems.