Short term forecasts of meteorological parameters play an important role in many societal processes. Until recently, seasonal autoregressive integrated moving average models (SARIMA) have been used to make forecasts on meteorological time series data. This thesis deploys and evaluates three different neural network forecasting systems, based on long short term memory (LSTM) networks. One univariate LSTM model, one multivariate LSTM model that receives all input parameters, and one multivariate LSTM model that only received correlating inputs. Each forecasting system uses twelve different LSTM submodels to forecast the meteorological parameters at the measuring site, Hannover-Herrenhausen. Theforecasting systems are compared with the SARIMA approach and a simple seasonal naive as a baseline model. For the comparison, the root mean squared error and mean absolute scaled error were computed. The neural network based forecasting systems outperform the SARIMA model in every parameter, except precipitation. Using only correlating inputs improved just selected parameter performance. Notably, the optimal window size was analysed to be 24 hours for the networks. The test on a second dataset from the measuring site in Ruthe revealed that the neural forecasting systems possess the ability to generalize on unknowndata.