Machine learning techniques for time series analysis

DAS Ltd has a systems engineering consultancy with a focus on Energy, Defence and Government sectors. DAS is currently supporting a number of clients in industrial data analytics projects primarily concerned with improving plant operations or increasing operational efficiency. DAS is specifically interested in developing techniques to make better use of the vast amount of data derived from an industrial process to improve plant operations through being able to forecast plant conditions and predict adverse events. The main goal of this PhD project is to develop machine learning (ML)/artificial intelligence (AI) techniques to accurately forecast time series, with applications to a wide range of industrial problems.

An example of application problem is foam formation in a biogas plant. Foaming occurs approximately four times a year in a number of power plants, which are part of DAS' client base. This can lead to various issues, including tank failure, which has many other implications. If the company is able to predict when this can happen, a number of measures can be taken to slow down the process or limit its impact on the process. For this example, we would like the methods to be designed in such a way that a number of indicators can be used to help the power plant operators to accurately predict when foaming will occur at a given plant.

One of the main difficulties in ML/AI methods is the evaluation of hyperparameters. Bilevel optimization methods will be combined with support vectors machines techniques to design cutting edge algorithms for the estimation of these parameters. The methods to be developed in this project will be scalable, so that they can easily be adapted to a wide range of situations and problems. Some of the algorithms to be developed will be programmed and included in a DAS Analytics "toolkit", which is under development.