Applying Recurrent Neural Networks for Short-Term Surface Temperature Forecasts Making

Today neural networks (NN) are omnipresent and its using is continuously growing in different areas like automation and optimization, researching, health care, engineering etc. This is due to NN possibility to recognize difficult patterns and structures. There are feed-forward and recurrent neural networks (RNN) among various architectures. Former implies only one-directional neuron’s relations, later permits bidirectional neuron’s interactions which mean possible cycles occurring, similar with cyclic graph. Nowadays RNN are within the best NN architectures for time series predictions. This paper we applied one of RNN architectures, namely long short-term memory (LSTM) for surface temperature time series prediction purposes, producing only short-term forecasts, not more than three hours lead time. Data was sourced from automatic weather station (AWS) archive, located in the Saint-Petersburg. The results obtained from conducted experiments showed LSTM fairly good performance. We also induced an algorithm to train LSTM and tackle with required predictions deadlines.