Short-term Temperature Forecasts using Deep Learning – an Application to Data from Ulm, Germany

With the increasing importance of solar energy we need more precise forecasts of its amount in order to guarantee network stability. Forecasting models are normally based on solar radiation, the major impact factor, but also on temperature which impacts the level of efficiency of the solar modules. Since conventional weather models (like the Lorenz model) are extremely chaotic, data driven models are getting more popular. Among those models the field of deep learning, i.e. neural networks with multiple hidden layers, is becoming more relevant. Especially convolutional neural networks which do not have the need of a feature extraction stage are a solid alternative to classic approaches. One of the main reasons is that computation power for massive parallel computing (i.e. GPGPU computing) is increasing. Authors like Dong et al. (2018) or Xiaoyun et al. (2016) use a recurrent neural network to predict temperature, wind speed or radiation, for example. Especially long short-term memory (LSTM) networks are often used, due to the fact that recurrent networks generally show good results when dealing with time series data (Lopez et al., 2016) and LSTM cells, in particular, have the "ability to bridge very long time lags" (Hochreiter & Schmidhuber, 1997), which is crucial when dealing with seasonal data. To the best of the authors knowledge, nobody applied the combination of convolutional and LSTM layers in this context so far, which is what we do in this work in order to forecast temperature data. Therefore we use weather data from Ulm between 2015 and 2018.