This thesis has looked into the possibility of classifying radio environment scenarios based on data received in base stations. It was done in order to improve forecasting of electrical output in these base stations. Subsequences of time series data was clustered with the k-means method, using dynamic time warping as the similarity measure and dynamic time warping barycenter averaging to find cluster centers. The subsequences were then classified and the labels were fed to the prediction models. The LSTM architecture was used to predict the electrical output. Two different architectures were used where one trained one model on all data and used the labels from clustering as an additional feature. The other trained multiple models on the different clusters found in clustering. What was found was that the k-means method could separate the subsequences into different radio environment scenarios. The introduction of clustered data decreased the mean square error of the prediction models of both architectures compared to baseline models trained on unclustered data.