undefined brand logo
undefined banner 3

Improving Non-Intrusive Load Monitoring with Ensemble Learning

Introduction:

In today's world, energy conservation has become a critical concern due to the increasing demand for energy and the scarcity of fossil fuels. One of the most effective ways to improve energy management is by having a better load identification and monitoring system. To achieve this, electric utilities worldwide have identified the benefits of more detailed metering options that enable loads to be monitored at the individual level.

Problem:

Traditionally, Non-Intrusive Load Monitoring (NILM) systems were based on the use of distributed direct sensing, where sensors were attached to each appliance that needed to be monitored. However, this approach was disruptive to user activities and was difficult to maintain. In contrast, NILM systems based on a single point of measurement, such as the main meter, are much less disruptive and easier to maintain. However, these systems require sophisticated algorithms to disaggregate the power consumption of each device.

Solution:

In this study, we proposed a novel ensembling technique that improves the performance of NILM systems based on deep neural network models. Our technique involves training a single model and then selecting a set of model instances with similar losses and averaging their weights to create a new model. This process results in a model that is less erratic and produces more accurate predictions.

Methods:

The methodology used in our research involved training a single model for the task of load disaggregation. Even after the loss had plateaued, the model was allowed to train further. At various points during the training process, the model weights were saved, and the model instances were evaluated on the test data to determine their validation loss.

After trying multiple combinations of these model instances, we observed that a combination of model instances at specific points in the training process resulted in the best performance. We then averaged the weights of these chosen model instances elementwise, creating a new model with the same architecture but with the averaged weight values. To evaluate the effectiveness of our proposed ensembling technique, we used the UK-DALE dataset, which includes power consumption data from five households. We chose the fridge and kettle as the target appliances for our study, as they have unique and repeating power consumption patterns. The data from house 1 was used for training and the data from house 2 for validation and testing. easy-to-use, and secure system that can enhance the convenience and comfort of any home.

Our preprocessing pipeline transformed the raw input power sensor readings into a format more suitable for neural networks. The pipeline extracted the appliance and main power consumption data, discarded records where either the appliance or main power sensor was inactive, broke the time series data at gaps in sensor reading, and resampled the data to a desired length.

Results:

Our results show that our ensembling technique significantly improves the performance of NILM systems. We observed a noticeable improvement in every single metric after the averaging process is done. Specifically, for the kettle model, we obtained a large improvement in both MAE and EA metrics.

Conclusion:

Our proposed ensembling technique is perfectly suited for NILM systems as it requires only a single model to be trained. This makes it especially useful for high power-consuming commercial and industrial buildings that consume three-phase power. Our ensembling technique proves to be even more valuable in this case, as the number of neural networks that need to be trained will be larger. We believe that our ensembling technique can help reduce the time and computational cost of training neural networks for NILM systems, making them more practical and efficient. In conclusion, our research on the application of ensemble learning to Non-Intrusive Load Monitoring (NILM) systems has produced significant improvements in performance and accuracy. Our proposed ensembling technique requires only a single model to be trained, making it highly efficient and suitable for NILM systems, particularly in high power-consuming commercial and industrial buildings that consume three-phase power. We believe that our ensembling technique has the potential to greatly reduce the time and computational cost of training neural networks for NILM systems, making them more practical and efficient for energy management. Our company is committed to continuing our research on this topic and developing new and innovative solutions for energy conservation and management.