Energy risk analysis

How many solar panels are installed on domestic roofs? How might these affect energy supply next year? In satellite images with a resolution of less than 1m per pixel, solar panels are very small objects scattered in a complex context, such as the combination of roads, buildings, roofs, parks, fields, etc. Image segmentation using convolution neural networks (CNNs) is able to capture and use this information for the final prediction. Working with an investment firm who assess risk in energy markets, EPCC designed and deployed a solar-panel-hunting CNN using the PyTorch framework across both GPU and CPU clusters, which achieved a classification performance of 83.8% recall and 84.6% precision. GPUs are ideally suited to this kind of image classification (they are, after all, graphical processing units), but by parallelising the PyTorch algorithm using both shared memory and MPI approaches we achieved comparable time performance using 10 CPU cores (shared memory) and 25 CPU cores (distributed memory MPI).