Machine Learning Approach to Identify Environmentally Friendly Alternatives to SF6 for Electricity Networks

六氟化硫（SF6）是一种强效温室气体，需要在现代输配电网络中逐步淘汰。每年有超过8000吨的SF6排放到大气中，其中80%用于电力行业，因此必须尽快减少其在气体绝缘设备中的使用，以避免对气候变化的负面影响。然而，SF6具有很好的电绝缘性能，这使得寻找替代品具有挑战性。由于现有气体组合的数量很多，对所有混合物进行高压测试是不切实际的。在其他研究中，机器学习方法被用于确定气体的电学性质，但尚未研究混合物的生成。因此，该项目的目的是开发一套工具，在机器学习的帮助下缩小SF6替代品的搜索空间。基于气体的已知电气特性、模拟和/或实验数据，机器学习将能够找到匹配所需操作特性的潜在候选者。然后，可以在受控的实验室环境中进一步测试所识别的气体，以验证结果。机器学习还可以帮助微调已知气体的混合物，以获得最佳性能。这种方法将大大加快识别过程，并最大限度地缩短上市时间。拟议产出将通过模块化软件开发方法实现。首先，将回顾有关机器学习和其他计算方法的文献，以选择适用于气体性质预测和混合物生成的方法。其次，将收集具有感兴趣特性的化合物的数据集来训练和测试算法。随着数据集的到位，将开发基于其评估的操作和环境参数的气体混合物优化软件。最后，主要软件的结果将在高压实验室进行测试，以验证识别的气体混合物的适用性并微调算法。此外，将探索通过第一性原理计算（如密度泛函理论）获得的分子描述符预测单个化合物的性质。这将使一种大规模的计算筛选方法成为可能，该方法将过滤掉不可用的化合物并突出潜在的SF6替代候选物。

The use of sulphur hexafluoride (SF6) in modern transmission and distribution networks needs to be phased out as it is a highly potent greenhouse gas. Annually more than 8000 tonnes of SF6 are emitted into the atmosphere and 80% of it is used in the power industry, therefore its use in gas insulated equipment must be reduced as soon as possible to avoid the negative contribution to the climate change. However, SF6 has great electrical insulation properties which makes finding a substitution challenging. Due to a high number of existing gas combinations, it is impractical to perform high voltage testing on all mixtures. The machine learning approaches are known to be applied to determine the electrical properties of gases in other research, but the generation of mixtures has not been investigated yet. Hence, the intent of this project is to develop a set of tools that will narrow down the search space for SF6 alternatives with the aid of machine learning. Based on the known electrical properties of gases, their simulation and/or experimental data, machine learning will be able to find the potential candidates matching required operational properties. Then the identified gases can be further tested in a controlled laboratory environment to verify the results. Machine learning can also help fine tune a mixture of known gases for optimal performance. This approach will significantly speed up the identification process and minimise the time to market. The proposed outputs will be achieved through a modular approach to software development. Firstly, the literature about machine learning and other computational approaches will be reviewed in order to select applicable methods both for gas properties prediction and mixture generation. Secondly, a dataset of compounds with the properties of interest will be gathered to train and test the algorithms. With the dataset in place, the software for gas mixture optimisation based on their evaluated operational and environmental parameters will be developed. Lastly, the results of the main software will be tested in the HV laboratory to verify the suitability of identified gas mixtures and to fine-tune the algorithms. Besides that, the prediction of the individual compound properties from molecular descriptors obtained via first principles calculations such as density functional theory will be explored. This will make a large computational screening approach possible which will filter out unusable compounds and highlight potential SF6 replacement candidates.