A Novel Recurrent Convolutional Neural Network for Ocean and Weather Forecasting

Numerical weather prediction is a computationally expensive task that requires not only the numerical solution to a complex set of non-linear partial differential equations, but also the creation of a parameterization scheme to estimate sub-grid scale phenomenon. The proposed method is an alternative approach to developing a mesoscale meteorological model – a modified recurrent convolutional neural network that learns to simulate the solution to these equations. Along with an appropriate time integration scheme and learning algorithm, this method can be used to create multi-day forecasts for a large region. The learning method presented is an extended form of Backpropagation Through Time for a recurrent network with outputs that feed back through as inputs only after undergoing a fixed transformation. An initial implementation of this approach has been created that forecasts for 2,744 locations across the southeastern United States at 36 vertical levels of the atmosphere, and 119,000 locations across the Atlantic Ocean at 39 vertical levels. These models, called LM3 and LOM, forecast wind speed, temperature, geopotential height, and rainfall for weather forecasting and water current speed, temperature, and salinity for ocean forecasting. Experimental results show that the new approach is 3.6 times more efficient at forecasting the ocean and 16 times more efficient at forecasting the atmosphere. The new approach showed forecast skill by beating the accuracy of two models, persistence and climatology, and was more accurate than the Navy NCOM model on 16 of the first 17 layers of the ocean below the surface (2 meters to 70 meters) for forecasting salinity and 15 of the first 17 layers for forecasting temperature. The new approach was also more accurate than the RAP model at forecasting wind speed on 7 layers, specific humidity on 7 layers, relative humidity on 6 layers, and temperature on 3 layers, with competitive results elsewhere