In this paper, three main approaches (univariate, multivariate and multistep) for electricity consumption forecasting have been investigated. In fact, three major algorithms (XGBOOST, LSTM and SARIMA) have been evaluated in each approach with the main aim to figure out which one performs the best in forecasting electricity consumption. The motivation behind this work is to assess the forecasting accuracy and the computational time/complexity for an embedded forecasting and model training at the smart meter level. Moreover, we investigate the deployment of the most efficient model in our platform for an online electricity consumption forecasting. This solution will serve for deploying predictive control solutions for efficient energy management in buildings. As a proof of concept, an already existing public dataset has been used. These data were mainly collected thanks to the usage of already deployed sensors. These provide accurate data related to occupancy (e.g., presence) as well as contextual data (e.g., disaggregated electricity consumption of equipment). Experiments have been conducted and the results showed the effectiveness of these algorithms, used in each approach, for short-term electricity consumption forecasting. This has been proved by performance evaluation and error calculations. The obtained results mainly shed light on the challenging trade-off between embedded forecasting model training and processing for being deployed in smart meters for electricity consumption forecasting.

Load forecasting is considered to be the core of an efficient predictive energy management for buildings. In this context, the deployment of smart meters and sensors enabled continuous energy usage monitoring in modern buildings. This streaming data led to the development of a new Online Home Energy Management System (OHEMS). The aim of this study is to develop an advanced smart metering infrastructure for online power forecasting, using embedded hardware with low computing power and real time constraints. As a benchmark, we applied both online and offline load forecasting modes to assess three prediction approaches in terms of accuracy and computational time. A single Machine learning algorithm using Long Short-Term Memory (LSTM) and hybrid Machine learning algorithms (CNN-LSTM), and ensembles machine learning approaches including eXtreme Gradient Boosting Machine (XGBoost) and Random Forest (RF). Furthermore, a novel practical stacking method for Short-Term Load Forecasting (STLF) using a stacked generalization ensemble method, which combines XGBoost and RF methods has been proposed. In online mode, the proposed stacking model achieved the best forecasting performance with a sMAPE of 1.15%, followed by RF (1.22%), XGBoost (1.23%), CNN-LSTM (1.97%) and LSTM (2.20%). In offline mode, the CNN-LSTM model outperformed all other methods with a sMAPE of 1.01%, demonstrating the advantage of deep feature extraction and complete data availability in batch forecasting. Performance-based retraining was shown to be more effective than periodic retraining, which might still be useful in fog computing scenarios. In general, offline CNN-LSTM is preferable for scenarios demanding maximum accuracy, while the stacking model is more suitable for scalable, real-time online forecasting in constrained environments.