The project was developed to propose a solution to ESO to achieve the Next Zero target by 2050 for the GB-wide energy system. In particular, the project was about understanding where different storage technology (e.g., short-duration tech solution in this case) solutions are needed when there exists a shortfall of energy from renewable sources.
The objective was to identify Grid Supply Points (GSPs) suitable for installing short-duration storage technology solutions like battery by signal analysis and develop a probabilistic machine learning model to visualise those suitable GSPs in 2D latent space derived from their very high dimensional space.
6-8 hours of off-peak time interval has been considered sufficient for a recharge of short-duration storage like battery. The period is sensible because of the general trend we see across the UK of an overnight low demand followed by a peak in the morning or a low demand in the afternoon followed by a peak in the evening.
Steps:
1. Fourier analysis – Frequency analysis of each time series GSP data has been done by Fourier transform. It helps us to label or categorize GSPs for battery or no-battery based on its Fourier spectral density measure
2. GPLVM – The unsupervised non-parametric statistical and machine learning model such as Gaussian Process Latent Variable Model has been applied on the labelled GSP data to visualize them in 2D latent space
pyro library from the PyTorch framework
1. A Very High dimensional GSP data can be visualised in 2-dimensional latent space by the GPLVM
2. As we move along from left to right in the 2D latent space, the scope for short term storage (battery) installation lessens and the variance of the latent points increases
3. Usually, in the area of separation or coexistence of the points from two categories (battery or no-battery), the priority of storage installation should be the least because of their nearly close variance and thereby proximity to each other