Résumé : This paper describes a flexible approach to short term prediction of meteorological variables. In particular, we focus on the prediction of the solar irradiance one hour ahead, a task that has high practical value when optimizing solar energy resources. As D\'efi EGC 2018 provides us with time series data for multiple sensors (e.g. solar irradiance, temperature, hygrometry), recorded every minute for two years and 5 geographical sites from La R\'eunion island, we test the value of using recently observed data as input for prediction models, as well as the performance of models across sites. After describing our data cleaning and normalization process, we combine a variable selection step based on AutoRegressive Integrated Moving Average (ARIMA) models, to using general purpose regression techniques such as neural networks and regression trees.