Significant advancements have been performed in predicting short term weather. This task has been almost complete with the help of State of the Art Deep Neural Networks. In particular, Recurrent Networks with Long-Short term Memory Cells have been developed for increased accuracy in local weather prediction. Each model was trained on the local climate for Vransko Slovenia and Montpellier France using 30 years of observations with frequency of 1 hour intervals. The historical weather data were purchased from Meteo Blue, which provided aggregated local data. Several combinations of architectures and layers were tested for different control horizons. The main focus was on 24hr and 12hr predictions to be used by the controller of the network for long and short term production planning respectively.
The models were training using high performance CUDA computing on local hardware. They were built using open source software such as Python 3.5 along with libraries such as Numpy, Pandas, Scikit, Tensorflow, Keras. Training was a very demanding process that took more than a day for each single combination of hyper-parameters. Among the hyper-parameters tested were different optimisers such as Adam, SGD, Adadelta, etc, different learning rates, decays, weight initialisers, stopping criteria along with visualisations supported by Tensorboard. As inputs weather variables were used including, temperature, humidity, sea level pressure, cloud coverage, precipitation, etc.