Today all meteorological prediction models are based on huge grid based finite element method calculations. Large sets of fluid dynamics differential equations are solved iteratively and the results are used as initial conditions for the next step. This computationally extremely expensive and the predictive accuracy is limited as errors will multiply for each predictive time step.
Using a combination of 3D ConvNets, LSTMs and a new architecture we are calling "Convolutional Transducing Mixture Density Networks" we are building an end-to end deep learning system with the aim of radically improving prediction results and cutting computational costs by orders of magnitude. What needs several hours of computations on a $40 million supercomputer can be done in less than 0.1 seconds on a laptop.
The first proof of concept for the system is being developed as a practical application for forecasting the power output of wind turbine generators. For this we have partnered with meteorological institutes and a number of energy companies that own and operate wind farms in Sweden.