An Intelligent Energy Aware Approach(s) for Smart Appliances

Abstract

Energy consumption is rising rapidly due to population proliferation, urbanization, and industrialization. Residential electricity demand is increasing rapidly, constituting about a quarter of total energy consumption. A large amount of energy can not be accumulated or transported; therefore, energy supply should be synchronized with consumption. With the advancement in information communication technology and rapid escalation in population, IoT paradigm has been adopted in homes for managing and optimizing energy consumption of home appliances. Electricity demand prediction is one of the sustainable solutions to improve energy efficiency in real-world scenarios. The non-linear and nonstationary consumption patterns in residential buildings make electricity prediction more challenging. Initially, comprehensive literature has been studied to explore energy-aware prediction and optimization techniques. The energy-aware prediction models have been reviewed and classified into machine learning and deep learning algorithms. From the literature, it has been inferred that non-linear and non-stationary consumption patterns in residential buildings make electricity prediction more challenging. The prediction models can be integrated with optimization approaches to optimize the predictive performance in residential buildings. Therefore, there is a need to develop an intelligent energy prediction and optimization approach for IoT-based smart homes and buildings. A multi-step prediction approach based on decomposition, reconstruction, and prediction models has been proposed to address these issues. Firstly, cluster analysis has been conducted to identify seasonal consumption patterns. Secondly, an improved CEEMDAN method and autoencoder model have been deployed to remove irregular patterns, noise, and redundancy from electricity load time series. Finally, the LSTM model and Bidirectional LSTM have been trained to predict electricity consumption by considering historical, seasonal, and temporal data dependencies. Further, an optimization approach is proposed using genetic algorithm for hyperparameter tuning of each Bidirectional LSTM model developed for heterogeneous home appliances. The proposed strategy selects the best hyperparameter values, namely optimal window size, and the number of neuron units, to obtain realistic, accurate future predictions. The energy datasets of Canada and Germany have been taken to validate the predictive performance of the proposed prediction model. These datasets contain a set of heterogeneous home appliances. The performance of the proposed hybrid energy prediction model is assessed using RMSE, MAE, and MSE. Eventually, the performance evaluation of the proposed GA-BiLSTM model has been performed on a real-time dataset of residential buildings. Real-time electricity consumption of residential buildings has been gathered from Punjab State Power Corporation Limited( PSPCL), Punjab, India to predict daily electricity demand using the GA-BiLSTM model, and the model capability is assessed against other state-of-the-art machine learning models such as SVR, RF, RNN, and LSTM. The inclusion of climatic conditions and temporal features improved the predictive performance of LSTM model. By incorporating GA optimization technique with bidirectional LSTM, the prediction accuracy of the proposed GA-BiLSTM model is improved than LSTM model for both real-time electricity dataset and I-BLEND dataset. The performance comparison of the proposed hybrid GA-BiLSTM model is made with existing state-of-the-art prediction models. The proposed approach outperformed the other state-of-the-art models and achieved the lowest mean absolute error.