Short-term temperature forecasts using a convolutional neural network — An application to different weather stations in Germany
David Kreuzer, Michael Münz, Stephan Schlüter
Abstract
Local temperature forecasts for horizons up to 24 h are required in many applications. A common method to generate such forecasts is the Seasonal Autoregressive Integrated Moving Average (SARIMA) model or, much simpler, the naïve forecast. In this paper, we test whether deep neural networks are able to improve on the results from the above mentioned methods. In addition to univariate long short-term memory (LSTM) networks, we present an alternative method based on a 2D-convolutional LSTM (convLSTM) network. For benchmarking our approach we set up a case study using data from five different weather stations in Germany. The SARIMA model and the univariate LSTM network perform quite well in the first few hours, but are then outperformed by the multivariate LSTM network and our convolutional LSTM network for longer forecast horizons. Besides, both multivariate approaches show better performance when the temperature is changing in the course of the day. Overall, our presented approach based on a convolutional LSTM network performs best on all used test data sets.