Tuesday, January 5, 2021

Investigating the impact of data quality on the energy yield forecast using data mining techniques

In this paper, we analysed the impact of using optimum combination of input variables and low dimensional subspace on Photovoltaic (PV) production forecasting accuracy. We worked in collaboration with Prof. Mussetta from Politecnico di Milano.

The main contribution presented in the paper is divided in two parts:
  1. Optimum combination of input meteorological features using feature extraction technique
  2. Low dimensional subspace using dimensional reduction technique
We assess and compare two cases when forecasting models are fed with all the features with the case when low subspace of dataset is used as an input to the models.
The simulation results reveal that depending on the location under study and the regression methods, using less variables as input to the forecasting models are enough to generate nearly similar results without affecting the performance. However, it is necessary to conduct the tests under different climatic conditions so as to ensure the reliability of the results. 
The figures below show the results obtained applying Pearsons correlation and principal component analysis. Figure 1 represents strength of association between two variables. Figure 2 is the biplot representation of the input features contributing variance on principal components PC1 and PC2.

Fig.1 : Pearson correlation map

Fig.2 : Biplot representation 

ISGT Europe 2020 was held virtually and we recorded the presentation for the same. The presentation is available at this link.


To support reproducibility and validating the results we have released the dataset utilized in the work along with the codes. 

Github repository with used dataset and evaluation code

For more information please see the paper:

Ekanki Sharma, Marco Mussetta, and Wilfried Elmenreich. Investigating the impact of data quality on the energy yield forecast using data mining techniques. In Proceedings of the IEEE PES Innovative Smart Grid Technologies Europe (ISGT-Europe). IEEE, October 2020.

Monday, November 16, 2020

Investigating the Benefit of Time-Series Imaging for Load Disaggregation

In this paper, we investigate the benefits of time-series imaging in load disaggregation, as we augment the wide-spread sequence-to-sequence approach by a key element: an imaging block.  

A Recurrence Plot
The approach presented in this paper converts an input sequence to an image, which in turn serves as input to a modified version of a common Denoising Autoencoder architecture used in load disaggregation. Based on these input images, the Autoencoder estimates the power consumption of a particular appliance. 

The main contribution presented in this paper is a comparison study of three common imaging techniques: 

  • Gramian Angular Fields, 
  • Markov Transition Fields, 
  • Recurrence Plots.
Further, we assess the performance of our augmented networks by a comparison with two benchmarking implementations, one based on Markov Models and the other one being a common Denoising Autoencoder. ´

The outcome of our study reveals that in 19 of 24 cases, the considered augmentation techniques provide improved performance over the baseline implementation. Further, the findings presented in this paper indicate that the Gramian Angular Field could be better suited, though the Recurrence Plot was observed to be a viable alternative in some cases. 

Our paper is to appear at the 7th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BuildSys ’20):

Hafsa Bousbiat, Christoph Klemenjak, and Wilfried Elmenreich. 2020. Exploring Time Series Imaging for Load Disaggregation. In The 7th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BuildSys ’20), November 18–20, 2020, Virtual Event, Japan.

 We are looking forward to discussing our paper at BuildSys!

Tuesday, October 20, 2020

Evolving NILM to NIAD: Non-Intrusive Activity Detection

Almost all documented practical use cases of load disaggregation rely on the analysis of appliance operational times and their impact on the monthly electricity bill. However, load disaggregation bears promising potential for other use cases. Recognizing user activities without the need to set up a dedicated sensing infrastructure is one such application, given that many household activities involve the use of electrical appliances. State-of-the-art disaggregation algorithms only provide support for the recognition of one appliance at a time, however. 

In collaboration with Andreas Reinhardt from TU Clausthal, we thus take load disaggregation to the next level, and present to what extent it is applicable to monitor user activities involving multiple appliances (operating sequentially or in parallel) using this technique. For the evaluation of our Non-Intrusive Activity Detection (NIAD), we synthetically generate load signature data to model nine typical user activities, followed by an assessment to what extent they can be detected in aggregate electrical consumption data. Our results prove that state-of-the-art load disaggregation algorithms are also well-suited to identify user activities, at accuracy levels comparable to (but slightly below) the disaggregation of individual appliances.

Our paper is to appear at the 2nd ACM Workshop on Device-Free Human Sensing (DFHS'20):

Andreas Reinhardt and Christoph Klemenjak. 2020. Device-Free User Activity Detection using Non-Intrusive Load Monitoring: A Case Study. In The 2nd ACM Workshop on Device-Free Human Sensing (DFHS ’20), November 15, 2020, Virtual Event, Japan.

 We are happily looking forward to pitching the concept of NIAD to the community!

Tuesday, October 13, 2020

Stop! Exploring Bayesian Surprise for Load Disaggregation

In our latest paper, which is the result of the ongoing collaboration between our lab and SFU's Computational Sustainability Lab, we bring the concept of Bayesian Surprise to NILM. When has enough prior training been done? When has a NILM algorithm encountered new, unseen data? We apply the notion of Bayesian surprise to answer these important questions for both, supervised and unsupervised algorithms.

"Bayesian surprise quantifies how data affects natural or artificial observers, by measuring differences between posterior and prior beliefs of the observers" - ilab.usc.edu

Bayesian Surprise is measured in "wow"
We compare the performance of several NILM algorithms to establish a suggested threshold on two combined measures of surprise: postdictive surprise and transitional surprise.

We provide preliminary insights and clear evidence showing a point of diminishing returns for model performance with respect to dataset size, which can have implications for future model development, dataset acquisition, as well as aiding in model flexibility during deployment.

The paper is to appear at the 5th International Workshop on Non-Intrusive Load Monitoring (NILM'20): 

Richard Jones, Christoph Klemenjak, Stephen Makonin, and Ivan V. Bajić. 2020. Stop! Exploring Bayesian Surprise to Better Train NILM. In The 5th International Workshop on Non-Intrusive Load Monitoring (NILM ’20), November 18, 2020, Virtual Event, Japan.

An author's copy can be obtained from Christoph's personal website

We are looking forward to discussing this novel approach for NILM.

Monday, August 3, 2020

SYND - A Synthetic Energy Dataset

As with related Machine Learning problems, applications like Non-Intrusive Load Monitoring (NILM) require a sufficient amount of data to train and validate new approaches. With SynD, we present a synthetic energy dataset emulating the power consumption of residential buildings. The dataset is freely available and contains 180 days of synthetic power data on aggregate level and individual appliances.

The SynD dataset is based upon
measurements of real devices
SynD is the result of a custom simulation process that relies on power traces of real household appliances. During a measurement campaign in two Austrian households, we monitored 21 electrical household appliances. The main goal of the measurement campaign was to record representative power consumption patterns of those 21 appliances, where a each pattern is represented by the shape of the power consumption over time for a single operation.

Technical validation of SYND by comparing with other datasets
In contrast to datasets entirely based on measurement campaigns, such as our dataset GREEND, the SynD dataset is constructed from a simulation model utilizing the measured devices. This way, a synthetic but realistic power consumption dataset can be obtained. In a technical validation of the dataset we compared SynD with a number of measured datasets showing that SynD is well within the varaiation between mutual datasets.

Wilfried Elmenreich states “Usually I rely on measured data, but with the SYND dataset, we are among the first who created a convincing synthetic dataset.”

The full paper describing the dataset is available under an open access policy here:
Christoph Klemenjak, Christoph Kovatsch, Manuel Herold, and Wilfried Elmenreich. A synthetic energy dataset for non-intrusive load monitoring in households. Scientific Data, 7(1):1–17, 2020. (doi:10.6084/m9.figshare.11940324)
The SynD dataset can be obtained freely at the SynD Github Repository.

Wednesday, July 22, 2020

Adaptive Weighted Recurrence Graph blocks for event-based NILM

To this day, hyperparameter tuning remains a cumbersome task in Non-Intrusive Load Monitoring (NILM) research, as researchers and practitioners are forced to invest a considerable amount of time in this task.

This paper proposes adaptive weighted recurrence graph blocks (AWRG) for appliance feature representation in event-based NILM. An AWRG block can be combined with traditional deep neural network architectures such as Convolutional Neural Networks for appliance recognition. Our approach transforms one cycle per activation current into a weighted recurrence graph and treats the associated hyper-parameters as learn-able parameters.

We evaluate our technique on two energy datasets, the industrial dataset LILACD and the residential PLAID dataset. The outcome of our experiments shows that transforming current waveforms into weighted recurrence graphs provides a better feature representation and thus, improved classification results. It is concluded that our approach can guarantee uniqueness of appliance features, leading to enhanced generalisation abilities when compared to the widely researched V-I image features. Furthermore, we show that the initialisation parameters of the AWRG’s have a significant impact on the performance and training convergence.

We provide the implementation of AWRG on Github.  If you find this tool useful and use it (or parts of it), we ask you to cite the following work in your publications:

A. Faustine, L. Pereira and C. Klemenjak, "Adaptive Weighted Recurrence Graphs for Appliance Recognition in Non-Intrusive Load Monitoring," in IEEE Transactions on Smart Grid, doi: 10.1109/TSG.2020.3010621.

Learn more about the authors Anthony Faustine, Lucas Pereira and Christoph Klemenjak.

Thursday, July 9, 2020

Supporting the Grid with Dynamic Residential Load Scheduling

Demand response (DR) for smart grids intends to balance the required power demand with the available supply resources. This is especially important with an increased amount of renewable energy sources since for most of them the energy yield cannot be shifted in time. For example, photovoltaic systems will provide their peak power at noon while customers might want to use energy at different times.

Residential load scheduling systems provide a solution to this problem by incentivizing consumers to use energy at times where production is high while motivating lower energy consumption in times of peak power. The goals of such residential load scheduling systems are therefore manifold: to cut peak power, to follow supply and to reduce the overall energy cost for the customer.

In a study we investigated different dynamic residential load scheduling systems with respect to optimal scheduling of household appliances on the basis of an adaptive consumption level pricing scheme (ACLPS). The proposed load scheduling system encourages customers to manage their energy consumption within the allowable consumption allowance of the proposed DR pricing scheme to achieve lower energy bills.

Dynamic pricing scheme motivating customers
to avoid energy consumption in peak periods

Simulation results show that employing the proposed approach benefits the customers by reducing their energy bill and the utility companies by decreasing the peak load of the aggregated load demand. For a given case study, the proposed residential load scheduling system based on ACLPS allows customers to reduce their energy bills by up to 53% and to decrease the peak load by  35%.

The full results are available in the paper

H. T. Haider, O. H. See, and W. Elmenreich. Dynamic residential load scheduling based on adaptive consumption level pricing scheme. Electric Power Systems Research, 133:27–35, 2016.