Use of time Series

Hi-

I'm looking to do some regression between two or more time series . I'm wondering if i should use the variable in matlab "timeseries". I don't have a point in the same time for each time series . I fixed this issue by creating an uniform time vector. Then , for each of my time series , I create a new time series with the same data but the uniform vector . when I don't have any data for one time - I take the closest data. This is working find but it's slow. Is there any better solution ? I put below the slow code which create my fulltime series Thanks In advance

% Create an uniforme time series - with a %t of Nsec.
%time and data represent my non uniform time series

function [fulltime,sol] = CreationFullTimeSeries(time,data,Nsec,start_date,end_date,start_time,end_time)
t1= datetime(strcat(start_date,start_time),'InputFormat','yyyyMMddHHmmSS');
t2= datetime(strcat(end_date,end_time),'InputFormat','yyyyMMddHHmmSS');
fulltime = transpose(t1:seconds(Nsec):t2);
sol = zeros(length(fulltime),size(data,2));

clear('temp');

j=1;
for i=1:length(fulltime)

     while and(datetime(time(j,:))<fulltime(i,1) ,j<size(time,1))
         j=j+1;
     end

     sol(i,:)=data(j,:);
  end

Answer 1

[Brian Neiswander]

Using a simple linear interpolation should be the easiest method to accomplish this. Let's say you have some sample measurement data in vectors named "time" and "data":

%define the starting and ending times of your measurements
start_date = '20100101';  %sstring containing tarting date in yyyyMMdd
end_date   = '20100102';  %string containing ending date in yyyyMMdd
start_time = '120000';    %string containing starting time in HHmmSS
end_time   = '120000';    %string containing ending time in HHmmSS

%make a time vector for the measurement data
Nsample = 2000;
t1 = datetime(strcat(start_date, start_time), 'InputFormat', 'yyyyMMddHHmmSS');
t2 = datetime(strcat(end_date, end_time), 'InputFormat', 'yyyyMMddHHmmSS');
time = t1:seconds(Nsample):t2;

%make some sample measurement data using a sine wave
data = sin(2*pi*linspace(0,1,length(time)));

Now, let's create a regularly spaced time vector "timeInterp" that you want to interpolate the measured data onto:

%make interpolating time vector
Nsec = 750;
timeInterp = t1:seconds(Nsec):t2;

Note that both the "time" and "timeInterp" vectors contain date strings. These will need to be converted to numerical values before interpolating. This can be done using the "datenum" function:

%convert date string vectors to date numbers
timeNum = datenum(time);
timeInterpNum = datenum(timeInterp);

Now you can use the "fit" function to interpolate the data onto the regularly spaced time vector:

%make a fit object that will linearly inerpolate over the data set
myfit   = fit(timeNum(:), data(:), 'linearinterp');

%interpolate the data onto the regularly spaced time vector using the fit object
dataInterp = myfit(timeInterpNum);

Note that the "fit" function is a part of the Curve Fitting Toolbox. You can check the interpolated data against the original data in a plot:

%plot the data
figure;
plot(timeNum, data, 'x', 'markersize',10);
hold on;
plot(timeInterpNum, dataInterp, 'k.');
datetick('x');  %change the x-axis labels to date strings
legend({'Measured data','Interpolated data'});

This approach will be much faster than using "for" and "while" loops. For more information, check the following documentation pages:

http://www.mathworks.com/help/matlab/ref/datenum.html

http://www.mathworks.com/help/curvefit/fit.html

http://www.mathworks.com/help/matlab/ref/datetick.html

Answer 2

[Peter Perkins]

Cedric, while the timeseries class does not support datetime arrays, it does have methods to synchronize or resample time series to a common time vector. You might find it convenient to use timeseries, even if you subsequently pull the time and data back out.