Estimation and forecasting battery state of health in real-world conditions using a data-driven approach

内容：鉴于减少二氧化碳排放的动力越来越大，随着可再生间歇性发电能力在全球能源结构中所占份额的增加，能源储存已成为一个主导话题。从电网规模到移动的应用，电池已成为提供高效储能解决方案的最有前途的替代品之一。因此，解决电池老化带来的问题已经引起了相当大的兴趣-随着时间的推移，存储容量和供电能力的衰减可能是从电网存储到电动汽车等应用中采用电池的最重要经济障碍。随着年龄的增长，健康状况的恶化是一个通过特定的设计和化学依赖机制影响所有可充电电池的问题。 在简单的意义上，电池的健康状态可以被定义为相对于制造时的能量，在给定的时间点放电时它可以释放的总能量。对这一数量的估计和衡量已有大量的现有研究。然而，准确性需要侵入性的、耗时的测试，导致几乎所有的分析都是在受控的实验室条件下产生的数据上进行的。对于在实际生产环境中运行的电池，估计变得更加困难，因为只能推断健康状态，而不是根据运行条件下生成的数据进行测量。困难是由各种因素，包括传感器的不准确性和变化，放电电流，时间和temperature.Aim：本项目的目的是开发的方法，以提供一个在线状态的健康估计电池在现实世界的条件下运行，其次是预测其随着时间的推移，鉴于观察到的使用模式的演变。主要的挑战是构建电池行为的数学模型。这用于将未观察到的健康状态与观察到的变量联系起来，这些变量（仅）包括电池端子之间的电压、从电池汲取的电流、时间和温度。该模型的复杂性是关键-模型必须充分代表现实，因为其输出必须以可接受的准确度映射到观察到的变量。另一方面，模型必须足够简单，以适用于所有按相同规格制造的电池。大型数据集的可用性至关重要，并且通常是模型复杂性和验证准确性的限制因素，因为从数量不足的噪声数据中推断参数会导致模型本身的不确定性达到不可接受的水平。新奇：利用“大数据”方面允许应用最先进的机器学习技术来推断数据中潜在的复杂关系，同时保持模型的鲁棒性。将机器学习应用于从各种实际操作条件下的多个电池中提取的数据，以估计和预测电池的健康状态，这是迄今为止尚未探索的领域。此外，使用真实世界的使用模式而不是实验室的使用模式，它将突出后者的结果转化为真实的操作环境的程度。价值：可靠的健康状态估计和预测对电池运营商至关重要。这个问题影响了整个投资生命周期内电池的经济性，从折旧预期到运营的可变成本。对于电池运营商来说，最直接的问题是维护计划，以最大限度地减少停机时间，以及客户对使用模式的警报，特别是对电池健康有害的警报。最终，模型可以内置到自动化电池管理系统中，从而优化使用，实现经济价值最大化。这属于EPSRC能源主题下的福尔斯。

Context: Given the increasing drive to reduce CO2 emissions, energy storage has become a dominant topic with the emergence of a larger share of renewable, intermittent generating capacity in the global energy mix. From grid-scale to mobile applications, batteries have become one the most promising alternatives to deliver an efficient energy storage solution. Consequently, addressing the issues that come with the aging of batteries has received commensurate interest - the fade of both storage capacity and ability to provide power over time is perhaps the most significant economic hurdle for battery adoption in applications ranging from grid storage to electric vehicles. The degrading state of health over aging is a problem that affects all rechargeable batteries via specific design- and chemistry-dependent mechanisms. In a simple sense, a battery's state of health may be defined as total energy it can release upon discharge at a given point in time relative to that at the time of manufacture. Estimation and measurement of this quantity has a substantial body of existing research. However, accuracy necessitates invasive, time-consuming tests resulting in almost all analysis having been conducted on data generated in controlled laboratory conditions. For batteries operating in a real-world production environment, estimation becomes considerably more difficult as state of health can only be inferred, rather than measured from data generated under operating conditions. The difficulty is compounded by the noisiness of available data resulting from a variety of factors including sensor inaccuracy and variations in discharge current, time and temperature.Aim: The aim of this project is to develop methods to give an online state of health estimate for batteries operating in real-world conditions, followed by forecasting its evolution over time given the observed pattern of usage. The main challenge is the construction of a mathematical model of battery behaviour. This serves to tie the unobserved state of health to the observed variables, consisting (exclusively) of the voltage across the battery terminals, current drawn from the battery, time and temperature. Complexity of this model is key - the model has to adequately represent reality, in that its outputs have to map to the observed variables to an acceptable degree of accuracy. On the other hand, the model has to be simple enough to be applicable over all batteries manufactured to the same specification. Availability of large datasets is paramount and is often the limiting factor for model complexity and validation accuracy, as inferring parameters from an insufficient quantity of noisy data results in an unacceptable level of uncertainty around the model itself. Novelty:Leveraging the "big data" aspect allows for state-of-the-art machine learning techniques to be applied to infer potentially complex relationships in the data while maintaining model robustness. Applying machine learning to data extracted from multiple batteries across a broad range of real-world operating conditions to estimate and forecast battery state of health is an area unexplored thus far. Additionally, using real-world usage patterns as opposed to laboratory ones, it will highlight the extent to which results from the latter translate to the real operating environment.Value:Reliable state of health estimates and forecasts are critical to battery operators. This issue affects the economics of operating batteries over the whole life cycle of the investment, from the expectation of depreciation, to the variable cost of operation. Of immediate concern for a battery operator would be maintenance planning to minimise down time as well as customer alerts for usage patterns particularly detrimental to battery health. Ultimately, models may be built into automated battery management systems which can optimise usage to maximise economic value. This falls under the EPSRC Energy theme.