Learnings

Outcomes

Storm AI is planned to be integrated as a beta feature of SSEN’s Power Track tool, a customer facing outage notification and reporting tool that allows consumers to see reported outages (along with planned/unplanned status, estimated time to restoration, and the status of the repair work), as well as reporting outages or damage to SSEN equipment. It’s currently planned as a BETA product as we expand the training data to increase the accuracy of the predictions. This will result in a BaU deployment of Storm AI.

Lessons Learnt

Efforts based around the improvement of the model’s performance in prediction were hampered to some extent by a lack of available classified image data samples on which to train the model. Efforts were made to introduce dataset augmentation; however, these greatly increased computation time, code-base complexity, memory usage, and exacerbated existing dataset biases (i.e. an imbalance in samples of different data labels across the set) without substantial increases in model performance.

As such, the efficacy of the model to be deployed into Power Track was not quite where it could be with due to the limited set of training data. A larger set of training data would be highly beneficial. Fortunately, this issue has been planned for with the inclusion of the ability to save models and their respective states after training. This has the effect of allowing a trained model to be retrained on new data, without having to train on the full set every time, thus allowing the model to be updated as new data is labelled.

As a learning point, for any machine learning driven project going forward, there should be as much labelled data as possible to ensure reliable and useful results are delivered. It should be noted that allowing for the model to be continually retrained, new data can be fed in as it becomes available.

The inclusion of multi-label classification for the prediction of asset types, including the ability to identify multiple assets in one image, has provided was challenging. This is a more complex operation than identifying a single damaged asset within an image for the model and requires additional steps within the model’s operations to train. Due to the complications in performing multi-label classification within the existing model, there are still issues to be worked through in regard to its predictive performance. This does not affect the overall functionality of the model and will not block its initial integration with Power Track but will require refining prior to being able to provide wholly reliable predictions.