Convert Smart* Dataset to List of Events

A new home energy dataset can be used to create a list of energy consumption activities is Smart* dataset. The information can be found at this link: http://traces.cs.umass.edu/index.php/Smart/Smart.

They have two house: B, and C that measured the aggregate electrical data, and environmental data; and House A that measured the aggregated energy consumption as well as electrical data from individual energy meters. As we are only interested in the individual appliances, so House A could be a dataset.

In house A, basically they recorded all sub-meters in 3 months, started from 01 May 2012 to 31 July 2012.
There are 91 days. Each day contains a list of labels, timestamp, and real power consumption at that timestamp. The resolution is about 3 seconds.

I attempted to get the data of every single event/label during these 3 months. The aim is to convert these dataset so that it would look like REDD dataset. Then, I can apply similar methods to get the list of ON-OFF EVENTS.

The first task is to separate the timestamp, and power usage of each label to single file. Having done that, it showed 53 labels altogether. To this far, the data should be looked at REDD dataset. Then, I converted these files to list of ON-OFF EVENTS.

I manually check the output. Surprisingly, majority of labels doesn't have data. For some reason, there are only few data on some irrelevant labels such as basement-dehumidifier, basement-freezer, bedroom-ac, kitchen-coffeepot, kitchen-toaster, livingroom-tv, master-fan. It is roughly 10/53 labels that working fine.
They don't have the data for major appliances like washing machine, microwave,...

This is a shame, because I have had a hope for this dataset.
Now I know it is unused for me. Such a waste of time :(.

Convert REDD & FigureEnergy raw data

I have finished the code of the mixture model of EGH extension, the prediction performance of this model is better as I expected, so it's good. More tests are required to confirm how robust the algorithm would be.

There are two real dataset: i) REDD, and ii) FigureEnergy. For the previous algorithm, I have converted these raw dataset into some machine readable using Matlab. However, it is such a pain to reuse the code, and generate the new format type of inputs by navigating the Matlab's code. So, I have switched to Python, which is objected-oriented and much easy to navigate the objects as well as function.

In REDD raw dataset, there are 6 houses. Each house consists of a list of appliances, measured the energy consumption by sub-meters. The energy consumption of each appliance is recorded every 3 seconds. As I am interested in predicting events, so I need to convert the raw dataset into ON-OFF Events. The algorithm is performed as follows:

• Firstly, set a threshold of energy consumption (typically 55W), store all segments that the appliances turned on.
• Secondly, set a GAP allowance parameter for two consecutive segments. If the gap between two consecutive segments are greater than the GAP allowance, we connect these two segments altogether, and considered as one segment.
• Thirdly, select the NOISE removal parameter, then filter all the segment that the duration is less then the Noise parameter.

In FigureEnergy dataset, the list of events does come altogether as one list. So, the idea is to program in Python to separate the appliances for every users to every single files. This is only required time-devoted for coding.

At the moment, the problem with REDD, and FigureEnergy data is the number of days that the dataset can give is small, and not enough to test how robust the prediction algorithm would perform. The solution could be to sample more data from the empirical distribution for every single appliances.

EGH_Extension

In the past, I have used a model, called "Mixture model of EGHs" to predict the next activity ahead. The idea is to analyse the past occurrences of the target event that we would like to predict. By looking the preceding window of the past occurrences, we can create a single sequence of these preceding sequences. The statistical mixture model could be generate to summarise this training dataset, so we can use this model to predict whether or not the target event would occur in the next time slot, given the recent window sequence of events.
The original algorithm of mixture model of EGHs considers all the occurrences of the events in the past. By applying to the home energy management scenario, we have observed that the day of the week is one of the important factor for the homeowners to decide which appliances/events/activities they could use. For example, the activities are obviously different between weekdays and weekends. This is because the homeowner' behaviours can be impacted by others external factors such as local weather, working hours, social activities.
Thus, taking the weekday as an important factor to predict the future events/activities, we can extend EGH algorithm. Instead of computing all past occurrences of the target event (i.e., the event that we would like to predict), we only consider the past occurrences that occurred to the same weekday that we would want to predict.
For example, if we want to predict occurrences of "washing machine" on next Tuesday, we only analyse the past preceding sequences that had the "washing machine" running on Tuesday.
By doing so, we might catch the patterns between appliances/activities and improve the prediction.
As the human's routines are best described daily, and weekly. The size of preceding window should be 7 days (1 week).

I have finished the code. Debugging at the moment...

Empirical Analysis

Having applied two state of the art algorithm from two other domains, I found that these algorithms might not work well in my home energy management scenario and could be difficult to extend the models because I don't have a clear idea how the real energy consumption activities performed in the home. Thus, I decided to go back to the real data and do some empirical analysis on some devices with the hope that I could find something interesting.

I started with the REDD dataset, with a couple of devices such as dishwasher, oven, and microwave. I converted the raw data into a list of ON-OFF events. Each event has a set of parameters such start time, duration, energy consumption. For example, the performance of the dishwasher of one house in REDD data can be seen as follows:

So, we can draw a dishwasher's distribution, which is looked like as follows:

This distribution does tell us that the dishwasher is most used between the consecutive days.

I have read a paper which mentioned that human's behaviour is on weekly (or daily) routines, so it does have some degree of regularity and predictability. And future prediction must be based on the assumption of determinism, so that the future events can be dependent on past events. Following to this idea, I convert from one single-dimensional time series into two dimensional matrix of weekdays and the state of the week that the event is turned ON-OFF. The data can be seen as follows:

In this table, we can see quite clearly that the user does not use the dishwasher on Thursday, and Friday, and likely to use once between Monday, and Tuesday. Also, he is likely to use in the weekend. The pattern is quite apparent, and could be found more effectively in this way.

Furthermore, each event will has a profile of how they perform weekly, and will have some certain degree of dependency between other events. This points out some challenges:
- how can we build a model to detect the dependency between events, and the dependency between weekdays.
- the length of the REDD data set is approximately 36 days, which is too short to detect any patterns. So, I might need to simulate the large scale of energy consumption activities for the "virtual house". However, how to simulate the synthetic data that can replicate the real house is still unsure.

All the predictive model can be started from this point.

Modelling discussion

 I have tried to build a probability model of user's events using Markov process. It is a simple model to calculate the probability of the next events that would occur in the next time step, based on the historical data. It uses the Markov property in which the next state depends only on the current state and not on the sequence of events that preceded it. It assumes the current state has all information from all previous state. This model can generate some results, however I don't think it generates a great result in our problem scenario.

As the labels can be dependent, therefore calculating the next state might require the information of some previous states, not just the current state. For example, we have an event sequence for using Washing Machine (WM). I assume the current state is today, which the WM is used, and we want to calculate the chance that the user will use the WM tomorrow. By applying Markov property, to estimate the probability of using WM tomorrow, we only need the information of the current state, which is the probability of using WM today. This probability is calculated from the historical data.

In a real scenario, using WM at the next state can be dependent on some previous use of WM in the past. Therefore, using Markov process in our model will not give a good result in prediction. However, we can use the model with Markov property as a bench mark to compare with other models, which I will research and implement.

I have read a paper of "A Model for Temporal Dependencies in Event Streams". They try to solve a very similar problem that I am trying to solve. They introduce the Piecewise-Constant Conditional Intensity Model (PCIM) to model the types and timing of events and capture the dependencies of each type of event on events in the past through a set of piecewise-constant conditional intensity functions.  The model is very complicated, but I want to implement this model to FigureEnergy data. They apply Bayesian Network Learning in their model, so I will try to read and understand this area more.

Mathematical model for event prediction

Predicting user's activities is not an easy task. Specifically, the activities themselves are dependent to each other, and the information of activities are given in uncertain ways (i.e., it is hard to get fully correct information of events in reality).

Previously, I have tried Poisson processes to predict user's events, however it has been uncesseful model as the events are independents in Poisson process. Continuing researching for an appropirate model, I have had a good chat with Long about a model, still using Markov process. I think we came with an appealing model, which also works for dependent data. I am implementing the model with the data collected from the FigureEnergy system, however there is not enough information to conclude this model. For further testing, I will test it with a large appropriate data, which are released by MIT. Thanksfully the data can be obtained from Oli. I will try to implement this model asap for further analysis.

Then, I will have to define a scenario in a formal mathematical way, then an existing model will be chosen to implement as a benchmark for future comparison. In addition, the paper of "A model for Temporal Dependencies in Event Streams" will be needed to read and understand as well.

A quick update of my work.

In the last meeting, we discovered that the Poisson process is not  a good model to apply in our scenario. It considers events are independent, while they are dependent in our case. Therefore, I need to switch to another appealing model, which should be done before 01 January 2012.

The strategy is to looking for models which work with dependent data. In addition, I need to gain much knowledge about machine learning and other mathematical models, so I am able to judge and select the right one.

I kick off with the paper of "Unsupervised Disaggregation of Low Frequency Power Measurements", which has been given by Oli. The paper mainly discusses the effectiveness of several unsupervised disaggregation methods using the factorial hidden Markov model on low frequency power measurements collected in real homes. In their model, the states of appliances are the hidden variables, and the aggregate power load is the observation. Therefore, they chose variants of Hidden Markov Model (HMM). More specially, they extend a Conditional Factorial Hidden Semi-Markov Model (CFHSMM), which allows the model to consider the dependencies between appliances and the dependencies on additional features, that I think is quite relevant to our case. Then, they apply machine learning process to estimate the parameters from the observations, and the hidden variables (which is the states of the appliances). Specifically, they use Expectation-Maximization algorithm (EM) to estimate the parameters, then using Maximum Likelihood Estimation (MLE) to estimate the hidden states.

In our case, the events are possibly annotated by users. So, I think Hidden Semi-Markov models could be used. I will check more references paper on CFHSMM to see if there is any relevant existing models.

Furthermore, I have grasped some book to read with the hope to have a better overal view on machine learning models. I will check out this list:

- Chapter 9: Mixture Models and EM (423-455) (book " Pattern Recognition and Machine Learning" - Christopher M. Bishop).

- Chapter 6: Bayesian Learning (154-199) (book: "Machine Learning" - Tom M. Mitchell).