This work explores the process of predicting energy consumption in smart buildings based on the consumption of devices and appliances. Particularly, this work studies the process of data analysis and generation of prediction models of energy consumption in Smart Buildings. Specifically, this article defines a feature engineering approach to analyze the energy consumption variables of buildings. Thus, it presents a detailed analysis of the process to build prediction models based on time series, using real energy consumption data. According to this approach, the relationships between variables are analyzed, thanks to techniques such as Pearson and Spearman correlations and Multiple Linear Regression models. From the results obtained with these, an extraction of characteristics is carried out with the Principal Component Analysis (PCA) technique. On the other hand, the relationship of each variable with itself over time is analyzed, with techniques such as autocorrelation (simple and partial), and Autoregressive Integrated Moving Average (ARIMA) models, which help to determine the time window to generate prediction models. Finally, prediction models are generated using the Long Short-Term Memory (LSTM) neural network technique, taking into account that we are working with time series. This technique is useful for generating predictive models due to its architecture and long-term memory, which allow it to handle time series very well. The generation of prediction models is organized into three groups, differentiated by the variables that are considered as descriptors in each of them. Evaluation metrics, RMSE, MAPE, and R² are used. Finally, the results of LSTM are compared with other techniques in different datasets.