Mamta Arora

Today, a major area of concern is the huge consumption of energy worldwide. The modern buildings are considered the largest energy consumers. When the energy consumption pattern of buildings is analyzed, attention is diverted towards the “Heating, Ventilation, and Air Conditioning” plant, which is accountable for the maximum consumption of energy in the buildings. This research work proposes a Deep learning framework including variants of Artificial Neural Network (ANN), namely, Recurrent Neural Network (RNN), Long-Short Term Memory (LSTM), and Gated Recurrent Units (GRU) to analyze and predict the abnormal energy consumption of the HVAC plants in commercial buildings. Also, proficient pre-processing, exploratory data analysis, and hyperparameter tuning for the proposed models are the key factors of this research. The proposed models have exhibited significant improvement as compared to the original RNN, LSTM, and GRU models. Different models of GRU were created using different values for hyperparameters, and the results of the experiments performed on the dataset show that one of the models of GRU outperformed the others with the least Mean Squared Errors (0.006) and (0.005) for the two months April and August, respectively. The research reveals that GRU is a better option for time series data analysis and is useful in energy consumption prediction for HVAC plants.

This research aims to propose a more sophisticated clustering and community detection technique in complex social networks through the use of neural networks; autoencoder, in particular. In the past, methods for network analysis and community detection used several graph algorithms, but with the advancements in deep learning, autoencoders are used for learning node features. These node representations are learnt using a neural network-based autoencoder and then clustering algorithms like k-means, agglomerative and spectral clustering are performed. These algorithms are then improved by incorporating with the Louvain algorithm for community detection. The proposed method, named the Spectral Louvain Algorithm, offers several advantages: it saves the stage of feature extraction, is suitable for Call Detail Record (CDR) and Social Network Analysis (SNA), does not require model retraining for different scale networks, and can work in mesh scale networks. It has better accuracy and performance than former approaches, even getting 100% of NMI and ARS, 78% of modularity in Karate Club and 89% of NMI, 93% of ARS and 86% of modularity in Dolphin data set with the least conductance. This method is particularly useful in discovering new relations and structures in a system and it is very efficient.