1
$\begingroup$

I would like to do classification of multi-source energy (wind/solar/teg) repersented in a time series data. My questions are : 1- What are the most relevent feature that I should chose to do the classification (statistical ones (kurtoisis/means/ variance...) on each sliding window (for experimental purpose) or spectral ones (DWT/FFT)) and which feature selection/extraction method is the best in this case. 2-What is the best classification method should I chose?

Thank you

$\endgroup$
  • $\begingroup$ Could you explain what you mean by 'classification' of time series, it is not clear, do you mean to take each data point as it is generated in time and classify that data point or are you attempting to identify recurrent patterns (timeseries tagging). If you could explain, it will probably be easier to answer the question. $\endgroup$ – Ironluca Apr 12 '17 at 12:50
  • $\begingroup$ In fact, As a start, I have a time series of solar energy during one year and I would like to classify it of(Summer/winter) energy classes. So I fix a size of a sliding window (so it contains many data points; for example 24 points if the sensor gives me one data each hour) and I calculate my features on each window. Depending on the classification training, I can predict on which class (summer/winter energy), data points of the testing data belong to. I hope that's clear. $\endgroup$ – LSola Apr 12 '17 at 14:24
  • $\begingroup$ OK, I shall provide a pespective, please see the answer below. $\endgroup$ – Ironluca Apr 13 '17 at 6:08
0
$\begingroup$

OK, let us say you have 24 data points per day, in a year you would have 8700~ data points. As an initial analysis you would likek to classify the data points in 2 classes, 'summer energy', 'winter energy'. Notice I am refering to data points and not time series. For this analysis, on way is not to take a time series view but a collection of datapoints view. Once you do that, it becomes a unsupervised classification problem. You could use K-Mean, where the value of K is 2. You could also use neural network based models like Adaptive Resonance Theory networks. As for feature selection, take all the features that you receive from the sensor. It could additionally help to center and scale the data.

In my opinion, if you want to classify the data into 2 classes, timeseries analysis is not immediately required, you could employ the method mentioned above and see how the results are.

EDIT 1: As I undestand (and assume), you have 4 sensors placed at 4 points of a single entity, e.g. an office. Each sensor output per hour is one of your dimensions. Hour of the day is also your data dimension. As an example, if you have 4 sensors, let us call them A,B,C,D, then your data point will look like:

Datapoint 1::
Vector position 0: value of A sensor, 1.0
Vector position 1: value of B sensor, 1.5
Vector position 2: value of C sensor, 0.5
Vector position 3: value of D sensor, 3.5
Vector position 4: hour of day (24), 14

You will have 720 such data points for one month. You should apply any clustering algorithm on the data points.

Few suggested clustering algorithms (there are many more):
1) K-Mean with K value as 2
2) ANN based approach

EDIT 2:

You could explore the ANN approach, you could make a 3 layer feed-forward network with 4 inputs and 2 outputs. The inputs would be the data point I mantioned above, however, it would be best to normalize the values ( here are some techniques) between 0 and 1. The output in your case will be 2 neurons, one representing summer and another winter, the input vector would be the datapoint vector and the output a vector of length 2, eg [0,1] for winter ideal valu & [1,0] for summer ideal value.

In another approach you could use K- Nearest Neighbour, in this case data normalization may not be strictly (technicaly) required.

For ANN library, you could use R/Python/Java, reference to one Java library is here.

For KNN, there are many library options, one reference is here

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thank your @Ironluca for your answer. However, I think that computing k-means statistics on the original data yields worse results. besides, avaluating clustering algorithms is really hard and I need to do classification and evaluate it, since I will be needed for implementation purpose. $\endgroup$ – LSola Apr 13 '17 at 7:54
  • $\begingroup$ How many dimensions, do you have for the data, if it is very high you could employ some available dimensionality reduction methods like PCA or compressional autoencoders $\endgroup$ – Ironluca Apr 13 '17 at 8:34
  • $\begingroup$ I have almost 8700 data. I know about PCA but actually, it is not that my problem. In fact, I am searching a method to do Classification of solar energy distributions, and for classification we need features that I am searhing for. $\endgroup$ – LSola Apr 13 '17 at 13:22
  • $\begingroup$ Is it uni-variate data or multi dimensional? If it is multi dimensional then those are your features (probable features). I am not clear as to what you mean by 'features that I am searching for'. Could you post few sample of the data points, if possible, that might clear my confusion. $\endgroup$ – Ironluca Apr 13 '17 at 14:41
  • $\begingroup$ first,I would like to thank you for trying to help me.Here are some plots of timeseries I have link.Suppose that you have sensors in different place and each one gets a value of energy every hour.I have plotted for each sensor the observations it gives during one month (720 datapoints, one hour each.As you see energy don't have necesarly the same range depending on the position of the sensor).My purpose is to find a way to do classification between summer/winter solarenergy. what features do I need for classification(statistical/spectral)& do I need to 'clean' my data? $\endgroup$ – LSola Apr 13 '17 at 16:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.