Information on energy consumption of individual appliances is important for demand response management, appliance usage pattern, and determination of appliance health. This information may be obtained by installing expensive energy meters with individual appliances. However, the cost becomes too high, and this method is non-scalable for large number of appliances. Energy disaggregation is a widely used low-cost computational approach to determine the energy consumption of individual appliances from the aggregated smart meter readings. Feature extraction, machine learning, and optimization techniques are generally used to discern appliance energy consumption from the aggregated readings with varying degrees of estimation accuracy. In recent years, Non-intrusive load monitoring (NILM) field, also known as energy disaggregation, is also benefiting from advancements in deep learning techniques such as, long short-term memory (LSTM), generative adversarial networks (GAN), and convolutional neural networks (CNNs). These new techniques require lot of data to train models but provide significant improvements over classical load disaggregation techniques.

So, what is DenseNILM? ConvNet corresponds to convolutional neural networks in Deep Learning, which based on spatial information, tries to obtain best possible distribution of data. Increase in number of convolutional processes are computationally expensive and can cause the model to not learn the real possible scenario. This causes model to give less accurate results. In this work, we propose a small and fast model architecture to enhance predictive capability by utilizing dense connectivity among convolutional layers (hence, Dense ConvNet). The method utilizes the information from all the preceding layers of model, and by connecting it with all successive layers, retain the information crucial for prediction. Predictive capability of model has been increased with the addition of attention mechanism, enforcing model to learn more precisely and provide better predictions. We have validated the performance of our proposed architecture on UK Domestic Appliance-Level Electricity (UK-DALE) dataset. UK-DALE is widely used in NILM literature, containing aggregated and individual appliance readings from multiple houses. DenseNILM provides better results as compared to previous literature utilizing deep learning techniques, with very low evaluation metric error in prediction of energy utilization of appliances.