Machine Learning Group (MLG)
Welcome to the Machine Learning Group (MLG) at the Calculus Energy. We are a highly versatile and a proactive group of researchers working on supervised and unsupervised learning, with a focus on energy applications (building energy consumption forecasting, solar PV estimation). Our interests within machine learning spans theoretical foundations, automatic differentiation, derivative and black box optimization algorithms - applied to a variety of industrial applications.
The nature of our industrial projects range from studying fundamental concepts in applied Bayesian statistics, all the way to getting our algorithms to perform competitively against the state-of-the-art in big-data applications. We also work on a broad range of application domains, including but not limited to: energy modelling, material science, finance, social networks, and chemical simulations, just to name a few. If you are interested in finding out more about our R&D activities and our bespoke AI products, please get in touch with your profile and the application domain (i.e the problem definition) you are interested in.
Control Engineering and Optimization
Dynamics, feedback, stability, optimisation — these and other related concepts are core to control science and engineering and are at the heart of the smart energy grid. Indeed, one could say that it is only a partial exaggeration to conclude that the smart grid is, at its core, is a control and optimisation problem.
At Calculus Energy, we strongly believe that control engineering is a key enabling technology for the deployment and higher penetration of renewable energy systems. Solar and wind power require advanced control techniques for high-performance and reliable operation. On the other hand, better predictive models of renewable generation are needed at all levels (which our Numerical Analysis group works on), right from component to systems. Therefore, the management of generation and load is now a multi-variable, non-convex optimisation problem with numerous constraints, and decision and control schemes are needed that factor in this complexity. We apply optimal design and engineering systems operation methodology to energy systems (storage, generation, distribution, and smart devices), and risk analysis. As a natural extension to this, our control and optimisation group also work within the multidisciplinary fields of signal processing, statistics and machine learning as a method for fitting parametric models to observed data.
Application example: Neural Networks to Predict Fluid Dynamics Solutions
In computational fluid dynamics (and some other problems in engineering), it often takes days or weeks to simulate the aerodynamic/engineering behavior of designs such as jets, spacecraft, or gas turbine engines. One of the major challenges encountered in these fields is question of how to simulate such systems efficiently with sufficient level of accuracy. A popular approach is to build reduced order models via say the Krylov Subspace and project this to a lower dimensional space. Another approach is to simply rely on tiny or vast datasets and enable these to make predictions. An ideal approach would have the speed of a neural network inference model and the accuracy of a reduced order models. CE strives to achieve this fine balance.
We model the fluid dynamics (and structural acoustics for example) problems via a combination of surface fitting, neural networks (CNNs, Neural ODEs) and cluster based networks with multiple layers. Such networks require a specialized training/computing procedure with the function networks trained by a classification loss function, and the context networks trained successively by a regression loss function. For a theortical survey of such methods, we point to reader to this paper. At CEs MLG group, we focus on such radical, model free, data driven approaches.
Get in touch with us to see how you can run your simulations more efficiently: